CN117216553A - Pre-training method, adjusting method, recommending method and related products of recommending model - Google Patents

Abstract

This application discloses a pre-training method, adjustment method, recommendation method and related products for a recommendation model. First, obtain the single-source domain behavior sequence of the object, and obtain the multi-modal information corresponding to the content in the single-source domain behavior sequence. The multi-modal information includes at least two different modal information; at the beginning of model pre-training, process Obtain a multi-modal vector representation of the content in a multi-domain universal content representation space; during model pre-training, the recommendation model to be trained is based on the behavior vector representation and predicts the first same source triggered after the end content of the single-source domain behavior sequence triggered by the object. domain content, and finally a preliminary recommendation model is obtained. Combined with the above model pre-training process, it can be seen that in this application, at least two different modalities of information are used, so that the information of various modalities can complement each other to build a sufficient training data set, so that the model obtained after pre-training Can be more robust.

Description

Pre-training methods, adjustment methods, recommendation methods and related products of recommended models

Technical field

This application relates to the field of multi-field recommendation technology, and in particular to a pre-training method, adjustment method, recommendation method and related products for a recommendation model.

Background technique

In the field of multi-domain recommendation (MDR, multi-domain recommendation) technology, great progress has been made in the training and construction of recommendation models. The built recommendation model can be used to predict objects in order to recommend objects of interest to them. Content. However, most of the existing recommendation model training solutions use text information as a bridge across domains, while ignoring the important value of other modal information. For example, image information is the main information source of content in many fields. This results in some recommendation models trained in related technologies facing insufficient modal information in the training data set, which in turn leads to insufficient model robustness.

Therefore, how to improve the robustness of recommendation models has become an urgent technical issue in the current field that needs to be solved.

Contents of the invention

The embodiments of this application provide a pre-training method, adjustment method, recommendation method and related products for a recommendation model, aiming to improve the robustness of the recommendation model.

The first aspect of this application provides a pre-training method for a recommendation model, including:

Obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents from the same source domain, and the multiple contents are sorted from first to last according to the time triggered by the object;

Obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities;

The multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal vector representation of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. ;

According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the object in the single-source domain behavior sequence is obtained. The behavior vector representation of the source domain to which the source domain behavior sequence belongs;

The recommendation model to be trained predicts, based on the behavior vector representation, the first content of the same source domain triggered after the object triggers the end content of the single-source domain behavior sequence;

Iteratively adjust the parameters of the recommendation model to be trained according to the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single-source domain behavior sequence. , until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed.

The second aspect of this application provides a model adjustment method for adjusting the preliminary recommendation model obtained through pre-training in the first aspect to realize the migration of the preliminary recommendation model from the source domain to the target domain, including:

Obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields arrive first according to the time triggered by the target object. Post-sequencing; multiple fields involved in the multi-domain mixed flow behavior sequence include the target domain;

Based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence, through the preliminary recommendation model, the multi-domain general content representation space of each content in the multi-domain mixed flow behavior sequence is obtained. Modal vector representation;

According to the ordering of each content in the multi-domain mixed flow behavior sequence and the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space, the target object is obtained at the location of the multi-domain mixed flow behavior sequence. Multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence;

The preliminary recommendation model predicts, based on the multi-domain mixed flow behavior vector representation, the content of the first target domain triggered by the target object after triggering the end content of the multi-domain mixed flow behavior sequence;

Iteratively adjust the preliminary recommendation model according to the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. parameters until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model.

The third aspect of this application provides a recommendation method for recommendation through the target recommendation model obtained in the second aspect, including:

Obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least includes content belonging to the target domain, and each content in the historical behavior sequence is in order from first to last according to the time triggered by the object to be recommended. sort;

Based on the multi-modal information corresponding to each content in the historical behavior sequence, through the target recommendation model, the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space is obtained respectively;

According to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space, the corresponding position of the object to be recommended in the historical behavior sequence is obtained. historical behavior vector representation;

The target recommendation model predicts, based on the historical behavior vector representation, the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence;

The content of the first target domain predicted by the target recommendation model is recommended to the object to be recommended.

The fourth aspect of this application provides a pre-training device for recommended models, including:

The behavior sequence acquisition module is used to obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents of the same source domain, and the multiple contents are arranged from first to last according to the time triggered by the object. sort;

A multi-modal information acquisition module is used to obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities;

The information input determination module is used to use the multi-modal information corresponding to the content as the input of the recommendation model to be trained, and obtain the multi-domain universal content representation of the content based on the input multi-modal information by the recommendation model to be trained. Multimodal vector representation of space;

A behavior representation construction module, configured to obtain a representation based on the sorting of content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space. The object’s behavior vector representation in the source domain to which the single-source domain behavior sequence belongs;

A same source domain content prediction module, configured to use the recommendation model to be trained based on the behavior vector representation to predict the first content of the same source domain triggered after the object triggers the end content of the single source domain behavior sequence;

The preliminary recommendation model acquisition module is used to iteratively adjust based on the difference between the content of the first same source domain that is predicted to be triggered and the content of the first same source domain that is actually triggered by the object after the end content of the single-source domain behavior sequence. The parameters of the recommended model to be trained are used until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed.

The fifth aspect of this application provides a model adjustment device, including:

The mixed flow behavior sequence acquisition module is used to obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields are in accordance with the subject The triggering time of the target object is ordered from first to last; the target domain is included in the multiple fields involved in the multi-domain mixed flow behavior sequence;

A multi-modal vector representation obtaining module is configured to obtain each content in the multi-domain mixed flow behavior sequence through the preliminary recommendation model based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence. Multi-modal vector representation of content in a multi-domain universal content representation space;

Mixed flow behavior vector representation acquisition module, used for sorting each content in the multi-domain mixed flow behavior sequence, and multi-modality of each content in the multi-domain mixed flow behavior sequence in a multi-domain universal content representation space Vector representation, obtaining the multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence of the target object;

A target domain content prediction module, configured to use the preliminary recommendation model based on the multi-domain mixed flow behavior vector representation to predict the first trigger of the target object after triggering the end content of the multi-domain mixed flow behavior sequence. The content of the target domain;

A target recommendation model acquisition module is used to obtain a module based on the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. , iteratively adjust the parameters of the preliminary recommendation model until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model.

The sixth aspect of this application provides a recommended device, including:

The historical behavior sequence acquisition module is used to obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least contains content belonging to the target domain, and each content in the historical behavior sequence is subject to the to-be-recommended object. The time of object triggering is sorted from first to last;

The historical multi-modal vector representation acquisition module is used to obtain the multi-domain universal information of each content in the historical behavior sequence through the target recommendation model based on the multi-modal information corresponding to each content in the historical behavior sequence. Multimodal vector representation of content representation space;

A historical behavior vector representation acquisition module, configured to obtain the to-be-listed content according to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in a multi-domain universal content representation space. The historical behavior vector representation of the recommended object corresponding to the historical behavior sequence;

A historical target domain content prediction module, configured to use the target recommendation model based on the historical behavior vector representation to predict the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence. ;

A target domain content recommendation module is configured to recommend the first content of the target domain predicted by the target recommendation model to the object to be recommended.

A seventh aspect of this application provides a computer device, which includes a processor and a memory:

The memory is used to store a computer program and transmit the computer program to the processor;

The processor is configured to perform the steps of the pre-training method of the recommendation model provided in the first aspect according to the instructions in the computer program, or perform the steps of the model adjustment method provided in the second aspect, or perform the recommendation method provided in the third aspect. A step of.

An eighth aspect of the present application provides a computer-readable storage medium. The computer-readable storage medium is used to store a computer program. When the computer program is executed by a computer device, it implements the pre-training method of the recommendation model provided in the first aspect. Steps, or perform the steps of the model adjustment method provided by the second aspect, or perform the steps of the recommended method provided by the third aspect.

A ninth aspect of the present application provides a computer program product, including a computer program that, when executed by a computer device, implements the steps of the pre-training method for the recommendation model provided in the first aspect, or executes the model adjustment method provided in the second aspect steps, or steps to perform the recommended method provided by the third party.

It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

In the technical solution of this application, the single-source domain behavior sequence of the object is first obtained, and the multi-modal information corresponding to the content in the single-source domain behavior sequence is obtained, where the multi-modal information includes information of at least two different modalities; in model pre-processing At the beginning of training, the multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal representation of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. state vector representation; during the model pre-training, based on the ordering of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the object is obtained in a single The behavior vector representation of the source domain to which the source domain behavior sequence belongs; in this way, based on the behavior vector representation, the recommendation model to be trained predicts the first content of the same source domain triggered after the end content of the single source domain behavior sequence is triggered by the object, and finally according to the prediction The difference between the first triggered content of the same source domain and the first content of the same source domain actually triggered by the object after the end of the single-source domain behavior sequence, iteratively adjust the parameters of the recommended model to be trained until the adjusted model meets the predetermined Training cut-off conditions, the preliminary recommended model is obtained after pre-training. Combined with the above model pre-training process, it can be seen that in this application, at least two different modalities of information are used, so that the information of various modalities can complement each other to build a sufficient training data set, so that the model obtained after pre-training Can be more robust. In addition, the use of pre-training technology to learn content and behavior sequences can be more versatile and robust, and the preliminary recommendation model can be effectively migrated to facilitate subsequent fine-tuning of the preliminary recommendation model, thereby achieving the best performance in model prediction and recommendation.

Description of drawings

Figure 1 is a schematic diagram of multi-domain modal information in related technologies provided by embodiments of the present application;

Figure 2 is a scene architecture diagram of the pre-training method, adjustment method, and recommendation method of a recommendation model provided in the embodiment of this application;

Figure 3 is a flow chart of a pre-training method for recommended models in practical application scenarios provided by the embodiment of the present application;

Figure 4 is a flow chart of a pre-training method for a recommendation model provided by an embodiment of the present application;

Figure 5 is a schematic structural diagram of a recommendation model provided by an embodiment of the present application;

Figure 6 is a schematic structural diagram of a multi-modal content representation constructor of a recommendation model provided by an embodiment of the present application;

Figure 7 is a schematic structural diagram of a multi-domain mapper of a recommendation model provided by an embodiment of the present application;

Figure 8 is a schematic diagram of the processing of a recommendation model provided by the embodiment of the present application;

Figure 9 is a flow chart of a model adjustment method provided by an embodiment of the present application;

Figure 10 is a schematic diagram of a process of adjusting the model provided by the embodiment of the present application;

Figure 11 is an adjustment schematic diagram of an adjustment model provided by an embodiment of the present application;

Figure 12 is a flow chart of a recommendation method provided by an embodiment of the present application;

Figure 13 is a comparison chart of the gain effects of the recommended model of this solution and the related technology model in practical applications provided by the embodiment of the present application;

Figure 14 is a schematic structural diagram of a pre-training device for a recommendation model provided by an embodiment of the present application;

Figure 15 is a schematic structural diagram of a model adjustment device provided by an embodiment of the present application;

Figure 16 is a schematic structural diagram of a recommendation device provided by an embodiment of the present application;

Figure 17 is a schematic structural diagram of a server in an embodiment of the present application;

Figure 18 is a schematic structural diagram of a terminal device in an embodiment of the present application.

Detailed ways

The embodiments of the present application are described below with reference to the accompanying drawings.

First, several terms that may be involved in the following embodiments of this application are explained.

Multi-modality: In the field of artificial intelligence, it often refers to the collaboration of sensory information, such as images, text, and voice, to help artificial intelligence understand the external world more accurately.

In the field of multi-domain recommendation (MDR, multi-domain recommendation) technology, great progress has been made in the training and construction of recommendation models. The built recommendation model can be used to predict objects in order to recommend objects of interest to them. Content. However, most of the existing recommendation model training solutions use text information as a bridge across domains, while ignoring the important value of other modal information. For example, image information is the main information source of content in many fields.

Figure 1 is a schematic diagram of multi-domain modal information in related technologies provided by embodiments of the present application. As shown in Figure 1, it is obvious that in the field of "books", text modal information is more important than image modal information. However, with the vigorous development of micro-video and communication technology, image information and other visual modal information It has gradually become the main information source of content in many fields. For example, in the field of "arts and handicrafts", image modal information is more important than text modal information. In this case, if the model pre-training is performed solely on the information of one modality (that is, the information of the text modality), this will inevitably lead to the modal information in the training data set faced by some recommendation models trained in related technologies. Inadequate, resulting in insufficient model robustness. Therefore, how to improve the robustness of recommendation models has become an urgent technical issue in the current field that needs to be solved.

In view of the above problems, this application provides a pre-training method, adjustment method, recommendation method and related products for the recommendation model, with the purpose of improving the robustness of the recommendation model. In the technical solution provided by this application, the single-source domain behavior sequence of the object is first obtained, and the multi-modal information corresponding to the content in the single-source domain behavior sequence is obtained, where the multi-modal information includes information of at least two different modalities. ; At the beginning of model pre-training, the multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, so that the multi-domain universal content representation of the content can be obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. Multi-modal vector representation of the space; during model pre-training, based on the ordering of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, Obtain the behavior vector representation of the object in the source domain to which the single-source domain behavior sequence belongs; in this way, the recommendation model to be trained can predict the first content of the same source domain triggered after the object triggers the last content of the single-source domain behavior sequence based on the behavior vector representation. , and finally based on the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end of the single source domain behavior sequence, iteratively adjust the parameters of the recommendation model to be trained until the adjustment The model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after pre-training. Combined with the above model pre-training process, it can be seen that in this application, at least two different modalities of information are used, so that the information of various modalities can complement each other to build a sufficient training data set, so that the model obtained after pre-training Can be more robust. In addition, the use of pre-training technology to learn content and behavior sequences can be more versatile and robust, and the preliminary recommendation model can be effectively migrated to facilitate subsequent fine-tuning of the preliminary recommendation model, thereby achieving the best performance in model prediction and recommendation.

Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

Artificial intelligence technology is a comprehensive subject that covers a wide range of fields, including both hardware-level technology and software-level technology. Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

The pre-training method, model adjustment method and recommendation method of the recommendation model provided in this application mainly involve machine learning. Among them, Machine Learning (ML) is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers can simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, teaching learning and other technologies.

The execution subject of the pre-training method of the recommendation model provided by the embodiment of the present application may be a terminal device, for example, the single-source domain behavior sequence of an object is obtained on the terminal device. Or the execution subject of the model adjustment method provided by the embodiment of the present application may be a terminal device, for example, the multi-domain mixed flow behavior sequence of the target object is obtained on the terminal device. Or the execution subject of the recommendation method provided by the embodiment of the present application may be a terminal device, for example, the historical behavior sequence of the object to be recommended is obtained on the terminal device. As examples, terminal devices may specifically include but are not limited to mobile phones, desktop computers, tablet computers, notebook computers, PDAs, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. The execution subject of the pre-training method of the recommendation model provided by the embodiment of the present application can also be a server, that is, the single-source domain behavior sequence of the object can be obtained on the server. Or the execution subject of the model adjustment method provided by the embodiment of the present application may be a server, for example, obtaining the multi-domain mixed flow behavior sequence of the target object on the terminal device. Or the execution subject of the recommendation method provided by the embodiment of the present application may be a server, for example, obtaining the historical behavior sequence of the object to be recommended on the terminal device. The pre-training method or the model adjustment method or the recommendation method of the recommendation model provided by the embodiments of the present application can also be executed collaboratively by the terminal device and the server. Therefore, in the embodiments of this application, there is no limitation on the implementation entity that implements the technical solution of this application.

Figure 2 exemplarily shows the scene architecture diagram of the pre-training method, adjustment method, and recommendation method of a recommendation model. The figure includes servers and various forms of terminal equipment. The server shown in Figure 1 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. In addition, the server can also be a basic cloud that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Cloud server for computing services.

In order to facilitate understanding of the technical solutions provided by the embodiments of the present application, next, a pre-training method for a recommendation model provided by the embodiments of the present application will be introduced based on a practical application scenario.

Referring to Figure 3, Figure 3 is a schematic diagram of a pre-training method for a recommended model in an actual application scenario provided by an embodiment of the present application. In this actual application scenario, the processing device is a server 300 with a model training function.

First, the server 300 obtains the object's single-source domain behavior sequence and the multi-modal information corresponding to the content in the single-source domain behavior sequence. The multi-modal information includes at least two different modal information (such as A-modal information and B-modal information). modal information), a single source domain behavior sequence includes multiple contents of the same source domain (such as a content and b content, where a content corresponds to A modal information and b content corresponds to B modal information), and multiple contents are triggered by objects The times are sorted from first to last. In order to make the vector representation corresponding to the multi-modal information more versatile, the server 300 uses the multi-modal information (ie, A-modal information and B-modal information) as the input of the recommendation model to be trained to process it through the recommendation model to be trained. Obtain the multi-modal vector representation of the content in a multi-domain universal content representation space. Then, based on the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the multi-domain universal content representation space, the behavior vector representation of the object is obtained, so that the server 300 can make the recommendation model to be trained according to the behavior vector representation. The prediction object triggers the first content of the same source domain after the last content of the single-source domain behavior sequence. Finally, the server 300 iteratively adjusts the parameters of the recommendation model to be trained based on the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single source domain behavior sequence, until the adjustment The final model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after pre-training. Combined with the above model pre-training process, it can be seen that in this application, at least two different modalities of information are used, so that the information of various modalities can complement each other to build a sufficient training data set, so that the information obtained after pre-training The model can be more robust.

Figure 4 is a flow chart of a pre-training method for a recommendation model provided by an embodiment of the present application. The pre-training method of the recommended model shown in Figure 4 includes:

S401: Obtain the single-source domain behavior sequence of the object.

The single-source domain behavior sequence includes multiple contents from the same source domain, and the multiple contents are ordered from first to last according to the time when they are triggered by the object. For example: multiple contents include content a, content b and content c. The object triggering time of content a is earlier than the object triggering time of content b. The object triggering time of content b is earlier than the object triggering time of content c. , then the single-source domain behavior sequence of the obtained object includes {content a, content b, content c}. It should also be noted that since objects can trigger content on different platforms, this application is not limited to obtaining content generated after being triggered by objects only on the same platform, where the platform may include an e-commerce platform. This is not specifically limited. For example: Platform 1 pushes the information of content a to the object. When the user clicks on the information of content a in platform 1, platform 1 will automatically jump to platform 2 to trigger content a.

S402: Obtain multi-modal information corresponding to the content in the single-source domain behavior sequence.

The multi-modal information includes information of at least two different modalities. For example: the modal information can be the information of the A mode, and the modal information can also be the information of the B mode. It can be understood that the multiple contents include content a, content b, and content c. The multimodal information corresponding to content a may only include information of modality A, and the multimodal information corresponding to content a may also include only modality B. Information, the multi-modal information corresponding to content a may also include information of modality A and information of modality B, and the same applies to content b and content c. It should be noted that in actual applications, the actual multi-modal information corresponding to the content needs to be determined based on the acquired information. In this way, various modal information complement each other in this application, and the content can be expressed more comprehensively.

S403: Use the multi-modal information corresponding to the content as the input of the recommendation model to be trained, and obtain the multi-modal vector representation of the content in the multi-domain universal content representation space by processing the recommendation model to be trained on the basis of the input multi-modal information. .

The multi-domain universal content representation space includes a space that supports processing content from multiple source domains to obtain a universal modal vector representation. It can be understood that the multi-domain universal content space is not limited to processing content from the same source domain. , where the multi-domain general content space also includes processing the content of the target domain, so that in the subsequent model fine-tuning process, each content in the target domain can be processed through the multi-domain general content space to obtain a general multi-modal vector representation. , so that the vector representation corresponding to the multi-modal information can be more versatile, that is, the vector representation output through the content space through the multi-domain can be uniformly recognized. In this stage, the multi-modal information corresponding to the content is used as the input of the recommendation model to be trained to obtain a multi-modal vector representation of the content in a multi-domain universal content representation space.

S404: Based on the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, obtain the source domain to which the object belongs in the single-source domain behavior sequence. Behavior vector representation.

In this step, the behavior vector representation includes a vector representation obtained after sorting multi-modal vector representations. Specifically, the multi-modal vector representation of the processed content in a multi-domain universal content representation space is combined in a single source domain. Sort the content in the behavior sequence to obtain the behavior vector representation of the source domain to which the object belongs in the single-source domain behavior sequence.

S405: Based on the behavior vector representation of the recommendation model to be trained, predict the first content of the same source domain triggered after the object triggers the last content of the single-source domain behavior sequence.

In this step, the first content of the same source domain includes content under the same source domain except the content in the single-source domain behavior sequence. For example: the content in the single-source domain behavior sequence includes content a, content b, and content c. Content d is the content predicted by the recommendation model to be trained. Content d belongs to the same source domain as content a, content b, and content c. In this way, the behavior vector representation obtained by content processing under the single-source domain behavior sequence is used to predict the content that does not exist in the single-source domain behavior sequence, but belongs to the same source domain as the content in the single-source domain behavior sequence, that is, the content to be The recommendation model is trained to predict whether the next content that the subject will click on is content d through content a, content b, and content c.

S406: Based on the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single source domain behavior sequence, iteratively adjust the parameters of the recommendation model to be trained until the adjustment The model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after pre-training.

It can be understood that in this step, the training model to be recommended is made to learn the difference between the first content of the same source domain that is predicted to be triggered, and the first content of the same source domain that is actually triggered by the object after the end content of the single-source domain behavior sequence. , and iteratively adjust the parameters of the recommended model to be trained based on the difference until the adjusted model meets the pre-training cutoff conditions, and finally the preliminary recommended model is obtained after the pre-training is completed. In this way, in this application, at least two different modalities of information are used, so that the information of various modalities can complement each other to build a sufficient training data set, so that the model obtained after pre-training can be more robust.

It should be noted that the pre-training cutoff conditions include the first condition and the second condition, where the first condition is the condition regarding the prediction loss, and the second condition is the condition regarding the contrastive learning comprehensive loss. The first condition includes that the prediction loss is less than the first loss threshold; wherein, the prediction loss is based on the difference between the content of the first source domain triggered by prediction and the content of the first source domain actually triggered. This application passes the pre-training cutoff.

The second condition includes that the comprehensive loss of contrastive learning is less than the second loss threshold; where the comprehensive loss of contrastive learning is the loss related to the cross-domain sequence and content comparison learning task and the cross-domain sequence and sequence comparison learning task. In the cross-domain sequence and content comparison learning task, the first content of the same source domain actually triggered after the end content of the single-source domain behavior sequence is used as a positive example, and other single-source domain behavior sequences input into the model in the same batch are used as positive examples. Contents involving other source domains in the source domain behavior sequence are used as negative examples, which are used to enhance the fusion and adaptation of universal representations of different domains. In the cross-domain sequence and sequence comparison learning task, the missing data sequence corresponding to the single-source domain behavior sequence is used as a positive example, and the single-source domain behavior sequence input into the model in the same batch as the single-source domain behavior sequence is used as a negative example. . The data missing sequence is obtained by randomly discarding the content in the single-source domain behavior sequence, or the data missing sequence is obtained by randomly discarding one or more modal information corresponding to the content in the single-source domain behavior sequence. It can be understood that the text modality information corresponding to the content in the single-source domain behavior sequence can be discarded, or the image modality information corresponding to the content in the single-source domain behavior sequence can be discarded.

In an implementable implementation, the comparative learning comprehensive loss can use the loss function To express, the loss function/> Expressed as follows:

in, Characterize the first loss function,/> Characterizes the second loss function, and λ represents the control loss function/> The second loss function/> The hyperparameters of the weights can make the adjusted model satisfy the loss function, and then the preliminary recommended model can be obtained after the pre-training is completed.

Specifically, the first loss function and the second loss function/> Expressed as follows:

Among them, the first loss function Represents the loss for cross-domain sequence and content comparison learning tasks, the second loss function/> represents the loss of cross-domain sequence and sequence contrast learning tasks, T represents the total number of single-source domain behavior sequences input into the model in the same batch, τ represents the temperature hyperparameter, s _j represents the single-source domain behavior sequence to be predicted, v _j represents the positive examples in the cross-domain sequence and content comparison learning task, v _j′ represents the negative examples in the cross-domain sequence and content comparison learning task,/> represents the positive examples in the cross-domain sequence and sequence comparison learning task, and s _j′ represents the negative examples in the cross-domain sequence and sequence comparison learning task.

Next, we introduce the specific structure of the recommendation model to be trained. Figure 5 is a schematic structural diagram of a recommendation model provided by an embodiment of the present application, as shown in Figure 5 .

The recommendation model to be trained includes a content representation construction module, a behavior representation construction module and a prediction module. The input end and output end of the behavior representation construction module are respectively connected to the output end of the content representation construction module and the input end of the prediction module. Among them, the content representation construction module is responsible for completing the multi-modal vector representation processing task, which is specifically used to process the multi-modal vector representation of the content in a multi-domain universal content representation space based on the input multi-modal information; behavioral representation The construction module is responsible for the task of obtaining behavior vector representation, which is specifically used to sort the content in the single-source domain behavior sequence and to represent the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space. Obtain the behavior vector representation of the source domain to which the object belongs in the single-source domain behavior sequence; the prediction module is responsible for the content prediction task, which is specifically used to predict the first identical trigger after the object triggers the end content of the single-source domain behavior sequence based on the behavior vector representation. content of the source domain, and when pre-training cutoffs are not met, backpropagation adjusts the parameters of the behavioral representation building block and/or the content representation building block.

Further, the content representation construction module includes a multi-modal content representation constructor and a multi-domain mapper. Step S403 (use the multi-modal information corresponding to the content as the input of the recommendation model to be trained, and process the multi-modal information corresponding to the content on the basis of the input multi-modal information through the recommendation model to be trained to obtain the multi-modal vector of the content in the multi-domain universal content representation space represents) including the following steps SA1-SA2 (it should be noted that steps SA1-SA2 are not shown in the drawings):

SA1: Input the multimodal information corresponding to the content into the multimodal content representation constructor, and perform joint representation learning on the multimodal information through the multimodal content representation constructor to obtain the multimodal representation of the content in the source domain to which it belongs.

Next, step SA1 will be described with reference to FIG. 6 . Figure 6 is a schematic structural diagram of a multi-modal content representation constructor of a recommendation model provided by an embodiment of the present application, as shown in Figure 6 . The multimodal content representation constructor includes a visual and language joint representation model and an adaptation layer. The adaptation layer includes an image modality self-attention module, a text modality self-attention module, and an image-text cross-modal self-attention module, at least two of which Different modalities of information include image modal information and text modal information. In FIG. 6 , image modality information is represented by A1 and text modality information is represented by A2.

It can be understood that the image modality information A1 and text modality information A2 corresponding to the content are jointly input into the visual and language joint representation model, so that the image modality information A1 and text modality information A2 are compared through the visual and language joint representation model. Perform joint representation learning to obtain the first image modal vector representation B1 and the first text modal vector representation B2 of the content in the source domain. The visual and language joint representation model can include the VilBERT model, which is not specifically limited here. In practical applications, other models that can jointly represent vision and language can be used.

After obtaining the first image modality vector representation B1 and the first text modality vector representation B2, the first image modality vector representation B1 is learned through the image modality self-attention module in the adaptation layer, and the image modality self-attention is obtained The content output by the module is represented by the second image modal vector representation C1 in the corresponding source domain; the first image modal vector representation B1 and the first text modal vector representation are jointly learned through the image and text cross-modal self-attention module in the adaptation layer. B2, obtain the first cross-modal vector representation C1C2 of the content output by the image text cross-modal self-attention module in the source domain, and learn the first text modal vector representation through the text modal self-attention module in the adaptation layer B2, obtain the second text modality vector representation C2 of the content output by the text modality self-attention module in the source domain to which it belongs. In this way, when the multi-modal content representation constructor includes a visual and language joint representation model and an adaptation layer, the multi-modal representation of the content in the source domain includes the second image modal vector representation, the second text modal vector representation and the third A cross-modal vector representation.

In an implementable implementation, the multimodal content representation constructor includes a visual and language joint representation model. The specific implementation method is the same as the above process. The only difference is that in this implementation method, the cooperation of the adaptation layer is not required. Multimodal vector representations are also available. In this way, when the multimodal content representation constructor includes a visual and language joint representation model, the multimodal representation of the content in the source domain includes the first image modality vector representation and the first text modality vector representation. Specifically, the image modality information and text modality information corresponding to the content are jointly input into the visual and language joint representation model, so as to perform joint representation learning on the image modality information and text modality information through the visual and language joint representation model. Obtain the first image modal vector representation and the first text modal vector representation of the content in the corresponding source domain. It should be noted that, after obtaining the first image modal vector representation and the first text modal vector representation, they can be directly input into the multi-domain mapper of the content representation construction module, and the multi-modal vector representation obtained is the same as the second modal vector representation. The image modality vector representation, the second text modality vector representation, and the first cross-modality vector representation are input into the multi-domain mapper of the content representation construction module to obtain the same multi-modal vector representation.

In another implementable implementation, the following formula can be used to obtain the first image modality vector representation obtained after joint representation learning of image modality information and text modality information through a visual and language joint representation model. First text modal vector representation. The formula is expressed as follows (in this formula, the visual and language joint representation model used is the VilBERT model):

x _i , y _i =VilBERT([[IMG]; e ₁ ,..., e _j ; [CLS]; w ₁ ,..., w _c ])

Among them, x _i and y _i represent the first image modal vector representation and the first text modal vector representation respectively; [IMG]; e ₁ ,..., e _j represent the image modal information, and j represents the image modal information. The number of;[CLS];w ₁ ,...,w _c represents the text modal information, and c represents the number of text modal information.

SA2: Map the multi-modal representation to the multi-domain universal content representation space through the multi-domain mapper, and obtain the multi-modal vector representation of the content in the multi-domain universal content representation space.

Next, step SA2 will be described with reference to FIG. 7 . Figure 7 is a schematic structural diagram of a multi-domain mapper of a recommendation model provided by an embodiment of the present application, as shown in Figure 7 . Multi-domain mapper includes mapping layer, concatenation layer and multi-layer perceptron. First, through the multi-domain mapper, the second image modality vector representation C1 of the content in the corresponding source domain is mapped to the multi-domain universal content representation space, and the first mapping result E1 is obtained; the first cross-modal representation of the content in the corresponding source domain is obtained. The vector representation C1C2 is mapped to the multi-domain universal content representation space to obtain the third mapping result E1E2; and the second text modal vector representation C2 of the content in the corresponding source domain is mapped to the multi-domain universal content representation space to obtain the second mapping result E2 . Specifically, the mapping layer includes a whitening layer and a hybrid expert network layer. The whitening layer includes a first whitening module, a second whitening module and a third whitening module corresponding to image modality, text modality and image-text cross-modality respectively; hybrid The expert network layer includes a first hybrid expert network, a second hybrid expert network and a third hybrid expert network respectively corresponding to image modality, text modality and image-text cross-modality; the first hybrid expert network and the second hybrid expert network and the third hybrid expert network both adopt gating mechanisms for multiple fields involved in the multi-domain universal content representation space; multiple fields include the source domain to which the content belongs.

Further, in order to reduce the redundancy between parameters and eliminate interference parameter information, the second image modal vector representation C1 is whitened through the first whitening module to obtain the first whitening result D1; the third whitening module is used to whiten the second image modal vector representation C1. A cross-modal vector representation C1C2 is whitened to obtain the third whitening result D1D2; and the second text modal vector representation C2 is whitened through the second whitening module to obtain the second whitening result D2.

In an implementable implementation, the whitening result obtained after processing by the whitening layer can be obtained by the following formula. The formula is expressed as follows:

in, and/> Both represent the whitening results _obtained after processing by the whitening _{layer .} Set, W ₁ represents the weight. It should be noted that the trainable parameters include parameters for obtaining the whitening results corresponding to the image modality vector representation, parameters for obtaining the whitening results corresponding to the text modality vector representation, and parameters for obtaining the corresponding cross-modal vector representations. The parameters of the whitening result are like this, which helps to improve the robustness of modal learning.

In order to further make adaptive adjustments to parameters with large differences between parameters, the first whitening result D1 can be processed through the first mixing expert network through the gating mechanism to obtain the first mapping result E1; through the third mixing The expert network processes the third whitening result D1D1 through the gating mechanism to obtain the third mapping result E1E2; and the second hybrid expert network processes the second whitening result D2 through the gating mechanism to obtain the second mapping result E2. Finally, the first mapping result E1, the third mapping result E1E2 and the second mapping result E2 are sequentially spliced through the splicing layer to obtain the splicing result, and the multi-layer perceptron is used to perform dimensionality reduction processing on the splicing result to obtain universal content in multiple domains. A multimodal vector representation of representation space.

In another implementable implementation, the mapping result obtained after processing by the hybrid expert network layer can be obtained by the following formula. The formula is expressed as follows:

FFN(x)=(GeLU(xW ₁ +b ₁ ))W ₂ +b ₂

Among them, _vi represents the mapping result obtained after processing by the hybrid expert network layer, gk represents the weight, G represents the number of gates in the hybrid expert network layer, k represents the kth gate, It represents the output of the kth gate after the whitening result, and k takes the value in the total number of gates G; GeLU represents the activation function, W ₁ and W ₂ both represent the matrix corresponding to the random parameter, and b ₁ and b ₂ both represent The bias of the matrix corresponding to the random parameter. It should be noted that the gating mechanisms G in the hybrid expert network layer are independent of each other and can be preset in advance, and the number of their gating mechanisms can be greater than the number of layers in the hybrid expert network layer.

Further, the behavior representation building module includes an object behavior encoder. Step S404 (According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, obtain the source domain to which the object belongs in the single-source domain behavior sequence. (behavior vector representation) includes the following steps SB1-SB3 (it should be noted that steps SB1-SB3 are not shown in the drawings):

SB1: Based on the sorting of the content in the single-source domain behavior sequence, obtain the position information of the content in the single-source domain behavior sequence.

SB2: Based on the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, and the position information of the content in the single-source domain behavior sequence, obtain the content representation corresponding to the single-source domain behavior sequence. sequence.

Specifically, the position information of the content in the single-source domain behavior sequence is first encoded into a position vector representation. After this stage, the multi-modal vector representation of the same content in the multi-domain universal content representation space and the single-source domain behavior sequence are The position vector representations in are added to obtain the vector superposition result corresponding to the content. Finally, based on the vector superposition results corresponding to each content in the single-source domain behavior sequence, the content representation sequence corresponding to the single-source domain behavior sequence is completed.

SB3: Use the content representation sequence as the input of the object behavior encoder, and encode the content representation sequence through the object behavior encoder to obtain the behavior vector representation of the source domain to which the object's single-source domain behavior sequence belongs.

Next, steps SB1-SB2 will be described with reference to FIG. 8 . Figure 8 is a schematic diagram of processing of a recommendation model provided by the embodiment of the present application, as shown in Figure 8 . Figure 8 shows the content representation corresponding to each content in the single source domain sequence (that is, the vector superposition result in the above text), and shows that each content in the single source domain sequence includes content a, content b, and content c. . As can be seen from Figure 8, the position information of the vector superposition result of content a in the single source domain sequence is position 3, the position information of the vector superposition result of content b in the single source domain sequence is position 1, and the vector superposition result of content c is at position 1. The position information in the single-source domain sequence is position 2. At this stage, the vector superposition results of content a, content b and content c are constructed into a content representation sequence {1, 2, 3} and input into the object behavior encoder. In this way, the content representation sequence is encoded by the object behavior encoder, and the behavior vector representation of the source domain to which the single-source domain behavior sequence belongs to the object is obtained.

It should be noted that the object behavior encoder in this application may include a transformer architecture. The transformer architecture includes: a multi-head self-attention layer (represented by MHAttn(·)) and a point feedforward network (represented by FFN(·)). The point feedforward network includes a multi-layer perceptron activated by ReLu. In summary, after obtaining multi-modal information based on the single source domain sequence of the object, this application inputs the multi-modal information corresponding to the content into the recommendation model to be trained, so as to obtain multi-domain universal content through content representation construction module processing. Represents the multi-modal vector representation of the space. At this stage, the multi-modal vector representation is processed according to the behavioral representation construction module to obtain the behavioral vector representation. Finally, the prediction module is used to reversely adjust when the pre-training cutoff condition is not met. Content represents the building blocks and\or behavior represents the parameters of the building blocks to obtain a preliminary recommendation model. In this way, a preliminary recommendation model can be obtained by pre-training and learning the single-source domain sequence of the object. In order to make the preliminary recommendation model more versatile, the preliminary recommendation model can be migrated from the source domain to the target domain, so that the preliminary recommendation model can adjust model parameters in the target domain to obtain a more accurate recommendation model. Next, we introduce the model adjustment method for preliminary recommendation model adjustment.

Figure 9 is a flow chart of a model adjustment method provided by an embodiment of the present application. A model adjustment method as shown in Figure 9 includes:

S901: Obtain the multi-domain mixed flow behavior sequence of the target object.

The multi-domain mixed flow behavior sequence includes multiple contents in multiple domains, and the multiple contents in multiple domains are ordered from first to last according to the time triggered by the target object. It should be noted that the multiple domains involved in the multi-domain mixed flow behavior sequence include the target domain. This model adjustment method is used to adjust the preliminary recommendation model obtained through pre-training through the above process, so as to realize the transition of the preliminary recommendation model from the source domain to Migration of target domain.

S902: Based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence, obtain the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space through the preliminary recommendation model. .

It can be understood that the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence obtained in this step in the multi-domain universal content representation space is different from the above process obtaining the multi-domain representation of each content in the single-source domain behavior sequence. Multimodal vector representations in a universal content representation space are indistinguishable.

S903: According to the sorting of each content in the multi-domain mixed flow behavior sequence and the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space, obtain the multi-domain mixed flow behavior of the target object Multi-domain mixed flow behavior vector representation corresponding to the sequence.

It can be further understood that the multi-domain mixed flow behavior vector representation corresponding to the target object's multi-domain mixed flow behavior sequence obtained in this step is the same as the behavior vector representation of the source domain to which the object's single-source domain behavior sequence belongs in the above process.

S904: Based on the multi-domain mixed flow behavior vector representation, the preliminary recommendation model predicts the content of the first target domain triggered by the target object after triggering the end content of the multi-domain mixed flow behavior sequence.

Next, step S904 will be described with reference to FIG. 10 . Figure 10 is a schematic diagram of a process of adjusting a model provided by an embodiment of the present application, as shown in Figure 10 . What is shown in Figure 10 is the content representation corresponding to each content in the multi-domain mixed stream sequence (that is, the vector superposition result in the above article), and shows that the multi-domain includes domain A, domain B and domain C, where domain A Each content of includes content a and content b, and shows that the position information of the vector superposition result of content a in the domain A sequence is position 1, and the position information of the vector superposition result of content b in the domain A sequence is position 3; Each content in domain B includes content c, and the position information of the vector superposition result of content c in the sequence of domain B is position 2; each content in domain C includes content d, and the position information of content d is shown. The position information of the vector superposition result in the domain C sequence is position 4. In this way, the multi-domain mixed flow behavior vector obtained by the construction is expressed as {(content a) 1, (content c) 2, (content b) 3, (content d) 4}. At this stage, the multi-domain mixed flow behavior vector representation is input into the object behavior encoder. After the target object triggers the end content of the multi-domain mixed flow behavior sequence through the preliminary recommendation model, the content of the first target domain triggered is content n .

S905: Based on the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end of the multi-domain mixed flow behavior sequence, iteratively adjust the parameters of the preliminary recommendation model until the model is adjusted to meet the requirements Preset fine-tuning cutoff conditions, and end the adjustment to obtain the target recommendation model.

Next, step S905 will be described with reference to FIG. 11 . FIG. 11 is an adjustment schematic diagram of an adjustment model provided by an embodiment of the present application, as shown in FIG. 11 . In order to better capture the relationship between the behavior vector representations of each content, content IDs are added to the behavior vector representations of each content, and each content ID is different. What is shown in Figure 11 is that the content of the triggered target domain is predicted by combining the multi-domain mixed flow behavior vector representation with the ID of each content corresponding to the multi-domain mixed flow behavior vector representation. Figure 11 shows that the id corresponding to (content a)1 is id1, the id corresponding to (content c)2 is id2, the id corresponding to (content b)3 is id3, and the id corresponding to (content d)4 is id4. The target domain is domain D. In this way, after the target object triggers the end content of the multi-domain mixed flow behavior sequence through the preliminary recommendation model, the content of the first target domain triggered is content n, and then based on the actual triggered content of the first target domain e and content n difference, iteratively adjust the parameters of the preliminary recommendation model until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model. It should also be noted that in this application, the focus is on adjusting the parameters of the multi-domain mapper part of the preliminary recommendation model.

In an implementable implementation, the multi-domain mixed flow behavior vector representation and predicted content can be obtained through the following formula. The formula is expressed as follows:

Among them, p _j represents the position of the content in the object behavior encoder, and F ^l represents the multi-domain mixed flow behavior vector representation. As F ^l , F ^l+1 represents the content of the first target domain of the predicted trigger, where the value of l can be customized.

It should also be noted that in this application, the probability of the predicted content can also be obtained through the following formula. The formula is expressed as follows:

in, Represents the probability of predicting the content of the next moment,/> Characterizes the behavioral sequence at the previous t moment,/> Represents the content to be predicted at the next moment,/> The content id corresponding to the content to be predicted at the next moment.

In summary, after adjusting the preliminary recommendation model through multi-domain mixed flow sequences, a more accurate target recommendation model is obtained. This target recommendation model can be used to recommend content that is of most interest to the recommendation object. Next, we introduce the recommendation method using this target recommendation model to treat recommended content of recommended objects.

Figure 12 is a flow chart of a recommendation method provided by an embodiment of the present application. In a recommendation method as shown in Figure 12, the recommendation method uses the target recommendation model obtained through the above model adjustment process for recommendation, including:

S1201: Obtain the historical behavior sequence of the object to be recommended.

The historical behavior sequence at least contains content belonging to the target domain, and each content in the historical behavior sequence is ordered from first to last according to the time when the recommended object is triggered.

S1202: Based on the multi-modal information corresponding to each content in the historical behavior sequence, through the target recommendation model, obtain the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space.

It can be understood that the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space is obtained in this step, which is different from the multi-domain universal content representation of each content in the single-source domain behavior sequence obtained in the above process. Multimodal vector representations in space are no different.

S1203: According to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space, obtain the historical behavior vector representation corresponding to the historical behavior sequence of the object to be recommended.

It can be further understood that the historical behavior vector representation corresponding to the historical behavior sequence of the object to be recommended in this step is obtained, which is the same as the behavior vector representation of the source domain to which the object's behavior sequence in a single source domain belongs in the above process.

S1204: Based on the historical behavior vector representation, the target recommendation model predicts the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence.

S1205: Recommend the content of the first target domain predicted by the target recommendation model to the object to be recommended.

Table 1 is a schematic representation of the data set of the pre-training method, adjustment method, and recommendation method of a recommendation model provided by the embodiment of the present application, as shown in Table 1. Avg.n represents the average length of the behavior sequence, and Sparsity represents the data sparsity. The household, clothing, and office data sets include the same source domain data set, and the food, instrument, electronic product, art, and sports categories include Target domain data set. It should be noted that in this application, in order to form a sequence for each content of the object in chronological order, the leave-one-out method can be used to divide the data set.

Table 1

Table 2 is a comparison table of the results of the pre-training method, adjustment method, and recommendation method of a recommendation model provided by the embodiment of the present application, as shown in Table 2. Compared with existing solutions, the method proposed in this solution comprehensively considers the content text, content image and content ID in the input data, and the proposed method comprehensively considers the pre-training model in the transfer learning technology used. , domain adaptive, hybrid flow and cross-domain technologies. In this way, the model constructed through this solution can be more robust, versatile and robust.

Table 2

Table 3 is a comparison table of the results of the pre-training method, adjustment method, and recommendation method of another recommendation model provided by the embodiment of the present application, as shown in Table 3. In this application, recall and/or normalized loss cumulative gain (NDGG) are used to evaluate the performance of the solution. In addition to 5, the values of recall and gain can also take values of 10, 15, and 20. etc. The specific value can also be set according to actual needs. Compared with the existing scheme, this scheme has an average improvement of +2.90 to +14.49% in terms of recall and gain (+14.49% is the result obtained after setting the recall to a value of 20), which allows the cross-modal mapper to The learned multi-modal vector representation is more robust and informative.

table 3

Refer to Figure 13. Figure 13 is a comparison chart of the gain effects of the recommended model of this solution and the related technology model in practical applications provided by the embodiment of the present application. As shown in Figure 13, the line comparison chart uses gain (NDGG@10) as the benchmark, in which the ordinate of each line chart represents the gain result, and the abscissa represents the data loss rate. Specifically, the 1301 line chart in Figure 13 represents the art data set. Line segment 1 in the 1301 line chart represents the assumption that image data is lost in this solution, line segment 2 represents the assumption that text data is lost in this solution, and line segment 3 represents the text in the existing solution. Data loss; the 1302 line chart in Figure 13 represents the electronic product data set, in which line segment 4 in the 1302 line chart represents the assumption that image data is lost in this solution, line segment 5 represents the assumption that text data is lost in this solution, and line segment 6 represents the existing solution in Chinese This data is lost; the 1303 line chart in Figure 13 represents the food data set, in which line segment 7 in the 1303 line chart represents the assumption that image data is lost in this solution, line segment 8 represents the assumption that text data is lost in this solution, and line segment 9 represents the existing solution in Chinese This data is lost. It can be seen that the method proposed in this scheme is significantly better than the method in the existing scheme even when the image data or text data is lost. The method proposed in this scheme is more robust and has more information in practical applications. .

Based on the pre-training method of the recommendation model provided in the foregoing embodiments, this application also provides a pre-training device for the recommendation model. Description will be made below with reference to Figure 14. Figure 14 is a schematic structural diagram of a pre-training device for a recommendation model provided by an embodiment of the present application. The pre-training device of the recommended model shown in Figure 14 includes:

The behavior sequence acquisition module 1401 is used to obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents of the same source domain, and the multiple contents are in order of arrival according to the time triggered by the object. post sorting;

The multi-modal information acquisition module 1402 is used to obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities;

The information input determination module 1403 is used to use the multi-modal information corresponding to the content as the input of the recommendation model to be trained, and process the multi-domain general content of the content based on the input multi-modal information through the recommendation model to be trained. Multimodal vector representation of representation space;

The behavior representation construction module 1404 is used to sort the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, Obtain the behavior vector representation of the source domain to which the single-source domain behavior sequence belongs for the object;

The same source domain content prediction module 1405 is used to predict the content of the first same source domain triggered after the object triggers the end content of the single source domain behavior sequence based on the behavior vector representation by the recommendation model to be trained;

The preliminary recommendation model acquisition module 1406 is used to iterate based on the difference between the content of the first same source domain that is predicted to be triggered and the content of the first same source domain that is actually triggered by the object after the end content of the single-source domain behavior sequence. Adjust the parameters of the recommended model to be trained until the adjusted model meets the pre-training cutoff conditions, and a preliminary recommended model is obtained after the pre-training is completed.

Optionally, the recommendation model to be trained includes a content representation construction module, a behavior representation construction module and a prediction module;

The content representation construction module is used to process and obtain a multi-modal vector representation of the content in a multi-domain universal content representation space based on the input multi-modal information;

The behavior representation construction module 1404 is used to sort the content in the single-source domain behavior sequence and the multi-modal vectors of each content in the single-source domain behavior sequence in the multi-domain universal content representation space. Represent, obtain the behavior vector representation of the object in the source domain to which the single-source domain behavior sequence belongs;

The prediction module is used to predict, based on the behavior vector representation, the content of the first same source domain triggered by the object after triggering the end content of the single-source domain behavior sequence, and when the pre-training cutoff condition is not met. , back propagation adjusts the parameters of the behavioral representation building blocks and/or the content representation building blocks.

Optionally, the content representation construction module includes: a multi-modal content representation constructor and a multi-domain mapper, and the information input determination module 1403 includes:

A multimodal information learning unit is used to input multimodal information corresponding to the content into the multimodal content representation constructor, and jointly represent the multimodal information through the multimodal content representation constructor. Learn to obtain the multi-modal representation of the content in the source domain;

A multi-modal information mapping unit configured to map the multi-modal representation to the multi-domain universal content representation space through the multi-domain mapper to obtain a multi-modal vector of the content in the multi-domain universal content representation space. express.

Optionally, the multi-modal content representation constructor includes: a visual and language joint representation model and an adaptation layer. The adaptation layer includes an image modality self-attention module, a text modality self-attention module, and an image-text cross-modality module. Modal self-attention module, the multi-modal information learning unit includes:

A modal information joint learning unit is used to jointly input image modal information and text modal information corresponding to the content into the visual and language joint representation model, and use the visual and language joint representation model to Perform joint representation learning on the information and the text modality information to obtain the first image modality vector representation and the first text modality vector representation of the content in the source domain to which it belongs;

A modality vector representation learning unit, configured to learn the first image modality vector representation through the image modality self-attention module, and learn the first text modality vector representation through the text modality self-attention module. , and jointly learn the first image modality vector representation and the first text modality vector representation through the image and text cross-modal self-attention module, to obtain the image modality self-attention module, the text The content output by the modal self-attention module and the image-text cross-modal self-attention module respectively is represented by the second image modal vector representation, the second text modal vector representation, and the first cross-modal vector representation of the source domain. .

Optionally, the multi-domain mapper includes a mapping layer, a splicing layer and a multi-layer perceptron, and the multi-modal information mapping unit includes:

A vector mapping result obtaining unit configured to map the second image modal vector representation, the second text modal vector representation, and the first cross-modal vector representation of the content in the corresponding source domain to the respective source domains through the multi-domain mapper. The multi-domain universal content representation space is used to obtain the first mapping result corresponding to the second image modal vector representation, the second mapping result corresponding to the second text modal vector representation and the first cross-modal vector representation. The corresponding third mapping result;

A mapping result splicing unit, configured to splice the first mapping result, the third mapping result and the second mapping result sequentially through the splicing layer to obtain a splicing result;

The splicing result dimensionality reduction processing unit is used to perform dimensionality reduction processing on the splicing result through the multi-layer perceptron to obtain a multi-modal vector representation of the content in the multi-domain universal content representation space.

Optionally, the mapping layer includes a whitening layer and a hybrid expert network layer. The whitening layer includes a first whitening module, a second whitening module and a third whitening module respectively corresponding to image modality, text modality and image-text cross-modality. Three whitening modules, the hybrid expert network layer includes a first hybrid expert network, a second hybrid expert network and a third hybrid expert network respectively corresponding to image modality, text modality and image-text cross-modality, the vector mapping Result acquisition units include:

The vector representation whitening processing unit is used to respectively represent the second image modality vector and the second text modality vector through the first whitening module, the second whitening module and the third whitening module. And perform whitening processing on the first cross-modal vector representation to obtain first whitening corresponding to the second image modal vector representation, the second text modal vector representation and the first cross-modal vector representation respectively. Result, second whitening result and third whitening result;

A whitening result gating processing unit configured to process the first whitening result, the first whitening result, and the third hybrid expert network through the gating mechanism through the first hybrid expert network, the second hybrid expert network, and the third hybrid expert network respectively. The second whitening result and the third whitening result are processed to obtain the first mapping result, the second mapping result and the third mapping result.

Optionally, the behavior represents the construction module, including:

A location information obtaining unit, configured to obtain the location information of the content in the single-source domain action sequence according to the sorting of the content in the single-source domain action sequence;

A content representation sequence acquisition unit, configured to obtain a multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, and the position of the content in the single-source domain behavior sequence. information to obtain the content representation sequence corresponding to the single-source domain behavior sequence;

A sequence encoding processing unit, configured to use the content representation sequence as an input to the object behavior encoder, and encode the content representation sequence through the object behavior encoder to obtain the object in the single source domain. Behavior vector representation of the source domain to which the behavior sequence belongs.

Optionally, the content represents the sequence acquisition unit, including:

A position vector representation encoding unit, used to encode the position information of the content in the single-source domain behavior sequence into a position vector representation;

A vector superposition result obtaining unit is used to add the multi-modal vector representation of the same content in the multi-domain universal content representation space and the position vector representation in the single-source domain behavior sequence to obtain the vector corresponding to the content. Overlay results;

A content representation sequence construction unit is configured to construct a content representation sequence corresponding to the single-source domain behavior sequence based on the vector superposition results corresponding to each content in the single-source domain behavior sequence.

Optionally, the device also includes a prediction loss adjustment module:

The prediction loss adjustment module is used for the pre-training cutoff conditions to include a first condition and a second condition, where the first condition is a condition regarding prediction loss, and the second condition is a condition regarding contrastive learning comprehensive loss; the third condition One condition includes: the prediction loss is less than the first loss threshold; wherein the prediction loss is obtained based on the difference between the content of the first source domain triggered by prediction and the content of the first same source domain actually triggered;

The second condition includes: the comprehensive loss of contrastive learning is less than the second loss threshold; wherein the comprehensive loss of contrastive learning is the loss related to the cross-domain sequence and content comparison learning task and the cross-domain sequence and sequence comparison learning task;

In the cross-domain sequence and content comparison learning task, the first content of the same source domain that is actually triggered after the end content of the single-source domain behavior sequence is used as a positive example and is input into the same batch as the single-source domain behavior sequence. Contents involving other source domains in other single-source domain behavior sequences in the model are used as negative examples;

In the cross-domain sequence and sequence comparison learning task, the data missing sequence corresponding to the single-source domain behavior sequence is used as a positive example, and the single sources of other source domains are input into the model in the same batch as the single-source domain behavior sequence. The domain behavior sequence is used as a negative example; the data missing sequence is obtained by randomly discarding the content in the single-source domain behavior sequence, or the data missing sequence is obtained by randomly discarding the content corresponding to the single-source domain behavior sequence. Obtained from one or more modal information.

Based on the model adjustment method provided in the foregoing embodiments, this application also provides a model adjustment device. Description will be made below with reference to Figure 15. Figure 15 is a schematic structural diagram of a model adjustment device provided by an embodiment of the present application. The model adjustment device shown in Figure 15 includes:

The mixed flow behavior sequence acquisition module 1501 is used to obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields are in accordance with the subject The time at which the target object is triggered is sorted from first to last; the multiple domains involved in the multi-domain mixed flow behavior sequence include the target domain;

The multi-modal vector representation obtaining module 1502 is configured to obtain the multi-domain mixed flow behavior sequence through the preliminary recommendation model based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence. Multi-modal vector representation of each content in a multi-domain universal content representation space;

The mixed flow behavior vector representation acquisition module 1503 is used for sorting each content in the multi-domain mixed flow behavior sequence and the multi-modality of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space. State vector representation, obtaining the multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence of the target object;

The target domain content prediction module 1504 is configured to use the preliminary recommendation model to predict, based on the multi-domain mixed flow behavior vector representation, the first trigger of the target object after triggering the end content of the multi-domain mixed flow behavior sequence. Describe the content of the target domain;

The target recommendation model acquisition module 1505 is configured to obtain the content of the first target domain actually triggered by the target object based on the predicted content of the target domain after the end content of the multi-domain mixed flow behavior sequence. difference, iteratively adjust the parameters of the preliminary recommendation model until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model.

Based on the recommendation method provided in the foregoing embodiments, this application also provides a recommendation device. Description will be made below with reference to Figure 16 . Figure 16 is a schematic structural diagram of a recommendation device provided by an embodiment of the present application. Recommended devices as shown in Figure 16 include:

The historical behavior sequence acquisition module 1601 is used to obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least contains content belonging to the target domain, and each content in the historical behavior sequence is subject to the to-be-recommended object. The time when the recommended objects are triggered is sorted from first to last;

The historical multi-modal vector representation acquisition module 1602 is used to obtain the multi-domain representation of each content in the historical behavior sequence through the target recommendation model based on the multi-modal information corresponding to each content in the historical behavior sequence. Multimodal vector representations of universal content representation spaces;

The historical behavior vector representation obtaining module 1603 is used to obtain the historical behavior vector representation according to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in a multi-domain universal content representation space. The historical behavior vector representation corresponding to the historical behavior sequence of the object to be recommended;

The historical target domain content prediction module 1604 is used to predict, based on the historical behavior vector representation by the target recommendation model, the first trigger of the target domain after the object to be recommended triggers the end content of the historical behavior sequence. content;

The target domain content recommendation module 1605 is configured to recommend the first content of the target domain predicted by the target recommendation model to the object to be recommended.

An embodiment of the present application provides a computer device, which may be a server. Figure 17 is a schematic structural diagram of a server provided by an embodiment of the present application. The server 900 may vary greatly due to different configurations or performance, and may include one or more central processing units (CPUs) 922 (for example, , one or more processors) and memory 932, one or more storage media 930 (eg, one or more mass storage devices) that stores applications 942 or data 944. Among them, the memory 932 and the storage medium 930 may be short-term storage or persistent storage. The program stored in the storage medium 930 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Furthermore, the central processor 922 may be configured to communicate with the storage medium 930 and execute a series of instruction operations in the storage medium 930 on the server 900 .

Server 900 may also include one or more power supplies 926, one or more wired or wireless network interfaces 950, one or more input and output interfaces 958, and/or, one or more operating systems 941.

Among them, CPU 922 is used to perform the following steps:

Obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents from the same source domain, and the multiple contents are sorted from first to last according to the time triggered by the object;

Obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities;

The multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal vector representation of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. ;

According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the object in the single-source domain behavior sequence is obtained. The behavior vector representation of the source domain to which the source domain behavior sequence belongs;

The recommendation model to be trained predicts, based on the behavior vector representation, the first content of the same source domain triggered after the object triggers the end content of the single-source domain behavior sequence;

Iteratively adjust the parameters of the recommendation model to be trained according to the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single-source domain behavior sequence. , until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed;

or;

Obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields arrive first according to the time triggered by the target object. Post-sequencing; multiple fields involved in the multi-domain mixed flow behavior sequence include the target domain;

Based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence, through the preliminary recommendation model, the multi-domain general content representation space of each content in the multi-domain mixed flow behavior sequence is obtained. Modal vector representation;

According to the ordering of each content in the multi-domain mixed flow behavior sequence and the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space, the target object is obtained at the location of the multi-domain mixed flow behavior sequence. Multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence;

The preliminary recommendation model predicts, based on the multi-domain mixed flow behavior vector representation, the content of the first target domain triggered by the target object after triggering the end content of the multi-domain mixed flow behavior sequence;

Iteratively adjust the preliminary recommendation model according to the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. parameters until the model is adjusted to meet the preset fine-tuning cutoff conditions, and the adjustment is completed to obtain the target recommendation model;

or;

Obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least includes content belonging to the target domain, and each content in the historical behavior sequence is in order from first to last according to the time triggered by the object to be recommended. sort;

Based on the multi-modal information corresponding to each content in the historical behavior sequence, through the target recommendation model, the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space is obtained respectively;

According to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space, the corresponding position of the object to be recommended in the historical behavior sequence is obtained. historical behavior vector representation;

The target recommendation model predicts, based on the historical behavior vector representation, the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence;

The content of the first target domain predicted by the target recommendation model is recommended to the object to be recommended.

An embodiment of the present application also provides another computer device, which may be a terminal device. As shown in FIG. 18 , for convenience of explanation, only the parts related to the embodiments of the present application are shown. If the specific technical details are not disclosed, please refer to the method part of the embodiments of the present application. Taking the terminal device as a mobile phone as an example:

Figure 18 shows a block diagram of a partial structure of a mobile phone provided by an embodiment of the present application. Referring to Figure 18, the mobile phone includes: radio frequency (English full name: Radio Frequency, English abbreviation: RF) circuit 1010, memory 1020, input unit 1030, display unit 1040, sensor 1050, audio circuit 1060, wireless fidelity (English full name: wirelessfidelity, English abbreviation: WiFi module 1070, processor 1080, power supply 1090 and other components. Those skilled in the art can understand that the structure of the mobile phone shown in FIG. 18 does not limit the mobile phone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange different components.

The following is a detailed introduction to each component of the mobile phone in conjunction with Figure 18:

The RF circuit 1010 can be used to receive and transmit information or signals during a call. In particular, after receiving downlink information from the base station, it is processed by the processor 1080; in addition, the designed uplink data is sent to the base station. Generally, the RF circuit 1010 includes but is not limited to an antenna, at least one amplifier, a transceiver, a coupler, a low-noise amplifier (English full name: LowNoise Amplifier, English abbreviation: LNA), a duplexer, etc. In addition, RF circuit 1010 can also communicate with networks and other devices through wireless communications. The above wireless communication can use any communication standard or protocol, including but not limited to Global System of Mobile communication (English full name: Global System of Mobile communication, English abbreviation: GSM), General Packet Radio Service (English full name: General Packet Radio Service, GPRS ), Code Division Multiple Access (English full name: CodeDivision Multiple Access, English abbreviation: CDMA), Wideband Code Division Multiple Access (English full name: Wideband CodeDivision Multiple Access, English abbreviation: WCDMA), Long Term Evolution (English full name: Long TermEvolution, English Abbreviation: LTE), email, short message service (full English name: Short Messaging Service, SMS), etc.

The memory 1020 can be used to store software programs and modules. The processor 1080 executes various functional applications and data processing of the mobile phone by running the software programs and modules stored in the memory 1020 . The memory 1020 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the storage data area may store a program according to Data created by the use of mobile phones (such as audio data, phone books, etc.), etc. In addition, the memory 1020 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device.

The input unit 1030 may be used to receive input numeric or character information, and generate key signal input related to user settings and function control of the mobile phone. Specifically, the input unit 1030 may include a touch panel 1031 and other input devices 1032. The touch panel 1031 , also known as a touch screen, can collect the user's touch operations on or near the touch panel 1031 (for example, the user uses a finger, stylus, or any other suitable object or accessory on or near the touch panel 1031 operation), and drive the corresponding connection device according to the preset program. Optionally, the touch panel 1031 may include two parts: a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch orientation, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into contact point coordinates, and then sends it to the touch controller. to the processor 1080, and can receive commands from the processor 1080 and execute them. In addition, the touch panel 1031 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 1031, the input unit 1030 may also include other input devices 1032. Specifically, other input devices 1032 may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control keys, switch keys, etc.), trackball, mouse, joystick, etc.

The display unit 1040 may be used to display information input by the user or information provided to the user as well as various menus of the mobile phone. The display unit 1040 may include a display panel 1041. Optionally, a liquid crystal display (English full name: Liquid Crystal Display, English abbreviation: LCD), organic light-emitting diode (English full name: Organic Light-Emitting Diode, English abbreviation: OLED), etc. may be used. to configure the display panel 1041. Further, the touch panel 1031 can cover the display panel 1041. When the touch panel 1031 detects a touch operation on or near it, it is sent to the processor 1080 to determine the type of the touch event. The processor 1080 then determines the type of the touch event. Type provides corresponding visual output on display panel 1041. Although in Figure 18, the touch panel 1031 and the display panel 1041 are used as two independent components to implement the input and input functions of the mobile phone, in some embodiments, the touch panel 1031 and the display panel 1041 can be integrated. Realize the input and output functions of mobile phone.

The mobile phone may also include at least one sensor 1050, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor. The ambient light sensor may adjust the brightness of the display panel 1041 according to the brightness of the ambient light. The proximity sensor may close the display panel 1041 and/or when the mobile phone is moved to the ear. or backlight. As a kind of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in various directions (usually three axes). It can detect the magnitude and direction of gravity when stationary. It can be used to identify applications of mobile phone posture (such as horizontal and vertical screen switching, related games, magnetometer attitude calibration), vibration recognition related functions (such as pedometer, tapping), etc.; as for other sensors such as gyroscope, barometer, hygrometer, thermometer, infrared sensor, etc. that can be configured on the mobile phone, we will not mention them here. Repeat.

The audio circuit 1060, speaker 1061, and microphone 1062 can provide an audio interface between the user and the mobile phone. The audio circuit 1060 can transmit the electrical signal converted from the received audio data to the speaker 1061, and the speaker 1061 converts it into a sound signal for output; on the other hand, the microphone 1062 converts the collected sound signal into an electrical signal, and the audio circuit 1060 After receiving, it is converted into audio data, and then processed by the audio data output processor 1080, and then sent to, for example, another mobile phone through the RF circuit 1010, or the audio data is output to the memory 1020 for further processing.

WiFi is a short-distance wireless transmission technology. The mobile phone can help users send and receive emails, browse web pages, and access streaming media through the WiFi module 1070. It provides users with wireless broadband Internet access. Although FIG. 18 shows the WiFi module 1070, it can be understood that it is not a necessary component of the mobile phone and can be omitted as needed without changing the essence of the invention.

The processor 1080 is the control center of the mobile phone, using various interfaces and lines to connect various parts of the entire mobile phone, and executing software programs and/or modules stored in the memory 1020 by running or executing them, and calling data stored in the memory 1020. Various functions of the mobile phone and processing data, thereby collecting overall data and information on the mobile phone. Optionally, the processor 1080 may include one or more processing units; preferably, the processor 1080 may integrate an application processor and a modem processor, where the application processor mainly processes operating systems, user interfaces, application programs, etc. , the modem processor mainly handles wireless communications. It can be understood that the above modem processor may not be integrated into the processor 1080.

The mobile phone also includes a power supply 1090 (such as a battery) that supplies power to various components. Preferably, the power supply can be logically connected to the processor 1080 through a power management system, so that functions such as charging, discharging, and power consumption management can be implemented through the power management system.

Although not shown, the mobile phone may also include a camera, a Bluetooth module, etc., which will not be described in detail here.

In this embodiment of the present application, the processor 1080 included in the mobile phone also has the following functions:

Obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents from the same source domain, and the multiple contents are sorted from first to last according to the time triggered by the object;

Obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities;

The multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal vector representation of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. ;

According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the object in the single-source domain behavior sequence is obtained. The behavior vector representation of the source domain to which the source domain behavior sequence belongs;

The recommendation model to be trained predicts, based on the behavior vector representation, the first content of the same source domain triggered after the object triggers the end content of the single-source domain behavior sequence;

Iteratively adjust the parameters of the recommendation model to be trained according to the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single-source domain behavior sequence. , until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed;

or;

Obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields arrive first according to the time triggered by the target object. Post-sequencing; multiple fields involved in the multi-domain mixed flow behavior sequence include the target domain;

Based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence, through the preliminary recommendation model, the multi-domain general content representation space of each content in the multi-domain mixed flow behavior sequence is obtained. Modal vector representation;

According to the ordering of each content in the multi-domain mixed flow behavior sequence and the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space, the target object is obtained at the location of the multi-domain mixed flow behavior sequence. Multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence;

The preliminary recommendation model predicts, based on the multi-domain mixed flow behavior vector representation, the content of the first target domain triggered by the target object after triggering the end content of the multi-domain mixed flow behavior sequence;

Iteratively adjust the preliminary recommendation model according to the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. parameters until the model is adjusted to meet the preset fine-tuning cutoff conditions, and the adjustment is completed to obtain the target recommendation model;

or;

Obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least includes content belonging to the target domain, and each content in the historical behavior sequence is in order from first to last according to the time triggered by the object to be recommended. sort;

Based on the multi-modal information corresponding to each content in the historical behavior sequence, through the target recommendation model, the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space is obtained respectively;

According to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space, the corresponding position of the object to be recommended in the historical behavior sequence is obtained. historical behavior vector representation;

The target recommendation model predicts, based on the historical behavior vector representation, the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence;

The content of the first target domain predicted by the target recommendation model is recommended to the object to be recommended.

Embodiments of the present application also provide a computer-readable storage medium for storing a computer program. When the computer program is run on a computer device, the computer device is used to execute the pre-training method of the recommendation model described in each of the foregoing embodiments. Any one of the implementations, or any one of the model adjustment methods described in the foregoing embodiments, or any one of the recommendation methods described in the foregoing embodiments.

Embodiments of the present application also provide a computer program product including a computer program, which, when run on a computer device, causes the computer device to execute any one of the pre-training methods for the recommendation model described in the foregoing embodiments, Either execute any one of the model adjustment methods described in the foregoing embodiments, or execute any one of the recommendation methods described in the foregoing embodiments.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be described again here.

In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only illustrative. For example, the division of the system is only a logical function division. In actual implementation, there may be other division methods. For example, multiple systems may be combined or integrated. to another system, or some features can be ignored, or not implemented. On the other hand, the coupling or direct coupling or communication connection between each other shown or discussed may be through some interfaces, and the indirect coupling or communication connection of the devices or units may be in electrical, mechanical or other forms.

The systems described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or software functional units.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (English full name: Read-OnlyMemory, English abbreviation: ROM), random access memory (English full name: Random Access Memory, English abbreviation: RAM), magnetic disk Or various media such as optical discs that can store computer programs.

The above embodiments are only used to illustrate the technical solutions of the present application, but are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments. Modifications may be made to the recorded technical solutions, or equivalent substitutions may be made to some of the technical features; however, these modifications or substitutions shall not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims (15)

1. A pre-training method for recommendation models, which is characterized by including: Obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents from the same source domain, and the multiple contents are sorted from first to last according to the time triggered by the object; Obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities; The multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal vector representation of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. ; According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the object in the single-source domain behavior sequence is obtained. The behavior vector representation of the source domain to which the source domain behavior sequence belongs; The recommendation model to be trained predicts, based on the behavior vector representation, the first content of the same source domain triggered after the object triggers the end content of the single-source domain behavior sequence; Iteratively adjust the parameters of the recommendation model to be trained according to the difference between the content of the first same source domain triggered by prediction and the content of the first same source domain actually triggered by the object after the end content of the single-source domain behavior sequence. , until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed. 2. The method according to claim 1, characterized in that the recommendation model to be trained includes a content representation construction module, a behavior representation construction module and a prediction module, wherein the input end and the output end of the behavior representation construction module are respectively connecting the output of the content representation construction module and the input of the prediction module; The content representation construction module is used to process and obtain a multi-modal vector representation of the content in a multi-domain universal content representation space based on the input multi-modal information; The behavior representation construction module is used to sort the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space. , obtain the behavior vector representation of the object in the source domain to which the single-source domain behavior sequence belongs; The prediction module is used to predict, based on the behavior vector representation, the content of the first same source domain triggered by the object after triggering the end content of the single-source domain behavior sequence, and when the pre-training cutoff condition is not met. , back propagation adjusts the parameters of the behavioral representation building blocks and/or the content representation building blocks. 3. The method according to claim 2, characterized in that the content representation construction module includes: a multi-modal content representation constructor and a multi-domain mapper; The multi-modal information corresponding to the content is used as the input of the recommendation model to be trained, and the multi-modal information of the content in the multi-domain universal content representation space is obtained by processing the recommendation model to be trained on the basis of the input multi-modal information. Vector representation, including: The multi-modal information corresponding to the content is input into the multi-modal content representation constructor, and the multi-modal information is jointly represented through the multi-modal content representation constructor to obtain the representation of the content in the source domain to which it belongs. multimodal representation; The multi-modal representation is mapped to the multi-domain universal content representation space by the multi-domain mapper, thereby obtaining a multi-modal vector representation of the content in the multi-domain universal content representation space. 4. The method according to claim 3, wherein the at least two different modalities of information include image modality information and text modality information; the multi-modal content representation constructor includes: visual and language A joint representation model and an adaptation layer, which includes an image modality self-attention module, a text modality self-attention module, and an image-text cross-modality self-attention module; The multi-modal information corresponding to the content is input into the multi-modal content representation constructor, and the multi-modal information is jointly represented through the multi-modal content representation constructor to obtain the content in the corresponding source. Multimodal representation of domains, including: The image modality information and text modality information corresponding to the content are jointly input into the visual and language joint representation model, and the image modality information and the text modality information are processed through the visual and language joint representation model. Joint representation learning to obtain the first image modal vector representation and the first text modal vector representation of the content in the source domain; The first image modality vector representation is learned through the image modality self-attention module, the first text modality vector representation is learned through the text modality self-attention module, and the image-text cross-modality The modal self-attention module jointly learns the first image modal vector representation and the first text modal vector representation to obtain the image modal self-attention module, the text modal self-attention module and the The content output by the image-text cross-modal self-attention module is the second image modal vector representation, the second text modal vector representation, and the first cross-modal vector representation of the source domain to which it belongs. 5. The method according to claim 4, wherein the multi-domain mapper includes a mapping layer, a splicing layer and a multi-layer perceptron; Mapping the multi-modal representation to the multi-domain universal content representation space through the multi-domain mapper to obtain a multi-modal vector representation of the content in the multi-domain universal content representation space includes: Through the multi-domain mapper, the second image modal vector representation, the second text modal vector representation and the first cross-modal vector representation of the content in the corresponding source domain are respectively mapped to the multi-domain universal content representation space, Obtain the first mapping result corresponding to the second image modality vector representation, the second mapping result corresponding to the second text modality vector representation, and the third mapping result corresponding to the first cross-modal vector representation; The first mapping result, the third mapping result and the second mapping result are sequentially spliced through the splicing layer to obtain a splicing result; The multi-layer perceptron performs dimensionality reduction processing on the splicing result to obtain a multi-modal vector representation of the content in the multi-domain universal content representation space. 6. The method according to claim 5, characterized in that the mapping layer includes a whitening layer and a hybrid expert network layer, and the whitening layer includes respectively corresponding to image modality, text modality and image-text cross-modality. A first whitening module, a second whitening module and a third whitening module; the hybrid expert network layer includes a first hybrid expert network and a second hybrid expert network corresponding to image modality, text modality and image-text cross-modality respectively. and a third hybrid expert network; the first hybrid expert network, the second hybrid expert network and the third hybrid expert network all adopt gates oriented to multiple fields involved in the multi-domain universal content representation space. control mechanism; the multiple domains include the source domain to which the content belongs; By using the multi-domain mapper, the second image modal vector representation, the second text modal vector representation and the first cross-modal vector representation of the content in the source domain are respectively mapped to the multi-domain universal content representation. space to obtain the first mapping result corresponding to the second image modality vector representation, the second mapping result corresponding to the second text modality vector representation, and the third mapping result corresponding to the first cross-modal vector representation. ,include: The second image modality vector representation, the second text modality vector representation and the first cross-modality are respectively processed through the first whitening module, the second whitening module and the third whitening module. The vector representation is whitened to obtain the first whitening result, the second whitening result and the third whitening result respectively corresponding to the second image modal vector representation, the second text modal vector representation and the first cross-modal vector representation. Three whitening results; Through the first hybrid expert network, the second hybrid expert network and the third hybrid expert network, the first whitening result, the second whitening result and the third whitening result are respectively processed through the gating mechanism. The whitening result is processed to obtain the first mapping result, the second mapping result and the third mapping result. 7. The method of claim 2, wherein the behavior representation construction module includes an object behavior encoder; According to the sorting of the content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, the location of the object is obtained. The behavior vector representation of the source domain to which the single-source domain behavior sequence belongs includes: According to the sorting of the content in the single-source domain behavior sequence, obtain the position information of the content in the single-source domain behavior sequence; According to the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, and the position information of the content in the single-source domain behavior sequence, the single-source domain is obtained The content corresponding to the behavior sequence represents the sequence; The content representation sequence is used as the input of the object behavior encoder, and the content representation sequence is encoded by the object behavior encoder to obtain the behavior of the object in the source domain to which the single source domain behavior sequence belongs. vector representation. 8. The method according to claim 7, characterized in that, according to the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space, and the location of the content in the multi-domain universal content representation space, Describe the position information in the single-source domain behavior sequence, and obtain the content representation sequence corresponding to the single-source domain behavior sequence, including: Encoding the position information of the content in the single-source domain behavior sequence into a position vector representation; Add the multi-modal vector representation of the same content in the multi-domain universal content representation space and the position vector representation in the single-source domain behavior sequence to obtain the vector superposition result corresponding to the content; According to the vector superposition results corresponding to each content in the single-source domain behavior sequence, a content representation sequence corresponding to the single-source domain behavior sequence is constructed. 9. The method according to any one of claims 1 to 5, characterized in that the pre-training cutoff condition includes a first condition and a second condition, wherein the first condition is a condition regarding prediction loss, and the second condition is Regarding the conditions for comparative learning comprehensive loss; the first condition includes: the prediction loss is less than the first loss threshold; wherein the prediction loss is based on the content of the first source domain triggered by prediction and the content of the first same source domain actually triggered gap gets; The second condition includes: the comprehensive loss of contrastive learning is less than the second loss threshold; wherein the comprehensive loss of contrastive learning is the loss related to the cross-domain sequence and content comparison learning task and the cross-domain sequence and sequence comparison learning task; In the cross-domain sequence and content comparison learning task, the first content of the same source domain that is actually triggered after the end content of the single-source domain behavior sequence is used as a positive example and is input into the same batch as the single-source domain behavior sequence. Contents involving other source domains in other single-source domain behavior sequences in the model are used as negative examples; In the cross-domain sequence and sequence comparison learning task, the data missing sequence corresponding to the single-source domain behavior sequence is used as a positive example, and the single sources of other source domains are input into the model in the same batch as the single-source domain behavior sequence. The domain behavior sequence is used as a negative example; the data missing sequence is obtained by randomly discarding the content in the single-source domain behavior sequence, or the data missing sequence is obtained by randomly discarding the content corresponding to the single-source domain behavior sequence. Obtained from one or more modal information. 10. A model adjustment method, characterized in that it is used to adjust the preliminary recommendation model obtained by pre-training by the method according to any one of claims 1 to 9, so as to realize the transition of the preliminary recommendation model from the source domain to the target domain. Migration; the model adjustment method includes: Obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields arrive first according to the time triggered by the target object. Post-sequencing; multiple fields involved in the multi-domain mixed flow behavior sequence include the target domain; Based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence, through the preliminary recommendation model, the multi-domain general content representation space of each content in the multi-domain mixed flow behavior sequence is obtained. Modal vector representation; According to the ordering of each content in the multi-domain mixed flow behavior sequence and the multi-modal vector representation of each content in the multi-domain mixed flow behavior sequence in the multi-domain universal content representation space, the target object is obtained at the location of the multi-domain mixed flow behavior sequence. Multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence; The preliminary recommendation model predicts, based on the multi-domain mixed flow behavior vector representation, the content of the first target domain triggered by the target object after triggering the end content of the multi-domain mixed flow behavior sequence; Iteratively adjust the preliminary recommendation model according to the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. parameters until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model. 11. A recommendation method, characterized in that the target recommendation model adjusted by the model adjustment method of claim 10 is used for recommendation; the recommendation method includes: Obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least includes content belonging to the target domain, and each content in the historical behavior sequence is in order from first to last according to the time triggered by the object to be recommended. sort; Based on the multi-modal information corresponding to each content in the historical behavior sequence, through the target recommendation model, the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space is obtained respectively; According to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in the multi-domain universal content representation space, the corresponding position of the object to be recommended in the historical behavior sequence is obtained. historical behavior vector representation; The target recommendation model predicts, based on the historical behavior vector representation, the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence; The content of the first target domain predicted by the target recommendation model is recommended to the object to be recommended. 12. A pre-training device for recommended models, characterized by including: The behavior sequence acquisition module is used to obtain the single-source domain behavior sequence of the object; the single-source domain behavior sequence includes multiple contents of the same source domain, and the multiple contents are arranged from first to last according to the time triggered by the object. sort; A multi-modal information acquisition module is used to obtain multi-modal information corresponding to the content in the single-source domain behavior sequence; the multi-modal information includes information of at least two different modalities; The information input determination module is used to use the multi-modal information corresponding to the content as the input of the recommendation model to be trained, and obtain the multi-domain universal content representation of the content based on the input multi-modal information by the recommendation model to be trained. Multimodal vector representation of space; A behavior representation construction module, configured to obtain a representation based on the sorting of content in the single-source domain behavior sequence and the multi-modal vector representation of each content in the single-source domain behavior sequence in the multi-domain universal content representation space. The object’s behavior vector representation in the source domain to which the single-source domain behavior sequence belongs; A same source domain content prediction module, configured to use the recommendation model to be trained based on the behavior vector representation to predict the first content of the same source domain triggered after the object triggers the end content of the single source domain behavior sequence; The preliminary recommendation model acquisition module is used to iteratively adjust based on the difference between the content of the first same source domain that is predicted to be triggered and the content of the first same source domain that is actually triggered by the object after the end content of the single-source domain behavior sequence. The parameters of the recommended model to be trained are used until the adjusted model meets the pre-training cutoff conditions, and the preliminary recommended model is obtained after the pre-training is completed. 13. A model adjustment device, characterized in that it includes: The mixed flow behavior sequence acquisition module is used to obtain the multi-domain mixed flow behavior sequence of the target object; the multi-domain mixed flow behavior sequence includes multiple contents in multiple fields, and the multiple contents in the multiple fields are in accordance with the subject The triggering time of the target object is ordered from first to last; the target domain is included in the multiple fields involved in the multi-domain mixed flow behavior sequence; A multi-modal vector representation obtaining module is configured to obtain each content in the multi-domain mixed flow behavior sequence through the preliminary recommendation model based on the multi-modal information corresponding to each content in the multi-domain mixed flow behavior sequence. Multi-modal vector representation of content in a multi-domain universal content representation space; Mixed flow behavior vector representation acquisition module, used for sorting each content in the multi-domain mixed flow behavior sequence, and multi-modality of each content in the multi-domain mixed flow behavior sequence in a multi-domain universal content representation space Vector representation, obtaining the multi-domain mixed flow behavior vector representation corresponding to the multi-domain mixed flow behavior sequence of the target object; A target domain content prediction module, configured to use the preliminary recommendation model based on the multi-domain mixed flow behavior vector representation to predict the first trigger of the target object after triggering the end content of the multi-domain mixed flow behavior sequence. The content of the target domain; A target recommendation model acquisition module is used to obtain a module based on the difference between the content of the target domain that is predicted to be triggered and the content of the first target domain that is actually triggered by the target object after the end content of the multi-domain mixed flow behavior sequence. , iteratively adjust the parameters of the preliminary recommendation model until the model is adjusted to meet the preset fine-tuning cutoff conditions, and then the adjustment is completed to obtain the target recommendation model. 14. A recommendation device, characterized in that it includes: The historical behavior sequence acquisition module is used to obtain the historical behavior sequence of the object to be recommended. The historical behavior sequence at least contains content belonging to the target domain, and each content in the historical behavior sequence is subject to the to-be-recommended object. The time of object triggering is sorted from first to last; The historical multi-modal vector representation acquisition module is used to obtain the multi-domain universal information of each content in the historical behavior sequence through the target recommendation model based on the multi-modal information corresponding to each content in the historical behavior sequence. Multimodal vector representation of content representation space; A historical behavior vector representation acquisition module, configured to obtain the to-be-listed content according to the sorting of each content in the historical behavior sequence and the multi-modal vector representation of each content in the historical behavior sequence in a multi-domain universal content representation space. The historical behavior vector representation of the recommended object corresponding to the historical behavior sequence; A historical target domain content prediction module, configured to use the target recommendation model based on the historical behavior vector representation to predict the content of the first target domain triggered after the object to be recommended triggers the end content of the historical behavior sequence. ; A target domain content recommendation module is configured to recommend the first content of the target domain predicted by the target recommendation model to the object to be recommended. 15. A computer device, characterized in that the device includes a processor and a memory: The memory is used to store a computer program and transmit the computer program to the processor; The processor is configured to execute the steps of the pre-training method of the recommendation model according to any one of claims 1 to 9 according to the instructions in the computer program, or to implement the steps of the model adjustment method of claim 10 when executed. , or when executed, the steps of the recommendation method described in claim 11 are implemented.