CN110232152A

CN110232152A - Content recommendation method, device, server and storage medium

Info

Publication number: CN110232152A
Application number: CN201910447811.9A
Authority: CN
Inventors: 林锡雄; 贺杰; 张文郁; 张磊; 张珩
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-05-27
Filing date: 2019-05-27
Publication date: 2019-09-13
Anticipated expiration: 2039-05-27
Also published as: CN110232152B

Abstract

The embodiment of the present application discloses a kind of content recommendation method, device, server and storage medium；The user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of the available target user of the embodiment of the present application, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship include the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended；Feature space mapping is carried out to user's characteristic information, obtains the corresponding user's characteristic information vector of user's characteristic information；Calculate the vector similarity in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector；Object content characteristic information vector is determined from content characteristic information vector based on vector similarity；The corresponding object content of object content characteristic information vector is chosen based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship；Recommend object content to target user.The embodiment of the present application can promote the accuracy of commending contents by obtaining content to be recommended similar with user's characteristic information as object content.

Description

Content recommendation method, device, server and storage medium

Technical field

This application involves Internet technical fields, and in particular to a kind of content recommendation method, device, server and storage Medium.

Background technique

In the information age, Infomiation Production and the consumption encouragement rapid development of information industry and information technology.Currently, it interconnects Net has become important information source, however, the swift and violent growth of the huge scale of internet and information resources brings letter The problem of breath overload, although that is, current information is resourceful, people are difficult effectively to obtain useful information.

It would generally recommend its interested content to user at present, but obtain the accuracy of user interest content at present It is low.

Summary of the invention

The embodiment of the present application provides a kind of content recommendation method, device, server and storage medium, can promote content The accuracy of recommendation.

The embodiment of the present application provides a kind of content recommendation method, comprising:

Obtain the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, the content DUAL PROBLEMS OF VECTOR MAPPING Set of relationship includes the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended；

Feature space mapping is carried out to the user's characteristic information, obtains the corresponding user characteristics of the user's characteristic information Information vector；

It calculates in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector Vector similarity；

Object content characteristic information vector is determined from content characteristic information vector based on the vector similarity；

The corresponding object content of object content characteristic information vector is chosen based on the content DUAL PROBLEMS OF VECTOR MAPPING set of relationship；

Recommend the object content to target user.

In some embodiments, the user's characteristic information includes user characteristics content and user characteristics vector, to described User's characteristic information carries out feature space mapping, obtains the corresponding user's characteristic information vector of the user's characteristic information, comprising:

By the user characteristics content map into default vector space, user characteristics content vector is obtained；

The user characteristics vector and user characteristics content vector are subjected to Vector Fusion, obtain the user's characteristic information Corresponding user's characteristic information vector.

In some embodiments, the user characteristics content map is obtained in user characteristics into default vector space Hold vector, comprising:

Word segmentation processing is carried out to the user characteristics content, obtains paragraph sequence, the paragraph sequence includes at least a kind of Paragraph；

The frequency of occurrences of the every kind of paragraph in the paragraph sequence is counted, the corresponding paragraph frequency of every kind of paragraph is obtained；

The paragraph that paragraph frequency meets preset threshold is chosen from the paragraph sequence, obtains target paragraph；

The corresponding user's characteristic information vector of the user's characteristic information is constructed according to the target paragraph.

In some embodiments, the user's characteristic information vector includes user's characteristic information high dimension vector, to the use Family characteristic information carries out feature space mapping, obtains the corresponding user's characteristic information vector of the user's characteristic information, comprising:

Dimension-reduction treatment is carried out to the user's characteristic information high dimension vector, obtains user's characteristic information low-dimensional vector；

The content characteristic information vector includes content characteristic information low-dimensional vector, calculates the user's characteristic information vector With the vector similarity in content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector, comprising:

Calculate content characteristic information low-dimensional vector in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Between vector similarity.

In some embodiments, dimension-reduction treatment is carried out to the user's characteristic information high dimension vector, obtains user characteristics letter Cease low-dimensional vector, comprising:

Using in Matching Model connection weight and biasing the user's characteristic information high dimension vector be weighted ask And processing, user's characteristic information high dimension vector after being handled, the Matching Model is by being labelled with the instruction of true vector similarity Practice sample training to form；

Dimensionality reduction is carried out to user's characteristic information high dimension vector after the processing using the default dimensionality reduction function in Matching Model Processing, obtains user's characteristic information low-dimensional vector.

In some embodiments, using in Matching Model connection weight and biasing to the user's characteristic information higher-dimension Vector is weighted summation process, user's characteristic information high dimension vector after being handled, comprising:

It is special to the user according to connection weight and biasing when the connection weight in the Matching Model is nonnegative number Reference breath high dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled；

When the connection weight in the Matching Model is negative, according to the absolute value of connection weight and biasing to described User's characteristic information high dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled.

In some embodiments, using in Matching Model connection weight and biasing to the user's characteristic information higher-dimension Vector is weighted summation process, after being handled before user's characteristic information high dimension vector, further includes:

Training sample is obtained, the training sample includes user's training sample, content training sample, the training sample mark The true vector similarity between user's training sample and content training sample is infused；

Initial matching model is trained using user's training sample, content training sample, obtains user's training Predicted vector similarity between sample and content training sample；

The initial matching model is restrained according to the true vector similarity and predicted vector similarity, is obtained Matching Model after training.

In some embodiments, the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user are obtained, is wrapped It includes:

Obtain the user's characteristic information of target user and the content characteristic information of content to be recommended, the content characteristic Information includes the feature of content to be recommended, content feature vector；

The feature of the content to be recommended is mapped in default vector space, is obtained in the feature of content to be recommended Hold vector；

The content feature vector and the feature vector of content to be recommended are subjected to Vector Fusion, obtained in be recommended The content characteristic information vector of appearance；

According to the mapping relations building between the content to be recommended and the content characteristic information vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

In some embodiments, the content characteristic information vector includes content characteristic information high dimension vector, will be described interior The feature vector for holding feature vector and content to be recommended carries out Vector Fusion, obtains the content characteristic information of content to be recommended After vector, further includes:

Dimension-reduction treatment is carried out to the content characteristic information high dimension vector, obtains content characteristic information low-dimensional vector；

According to the mapping relations building between the content to be recommended and the content characteristic information vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, comprising:

According to the mapping relations between the content to be recommended and the content characteristic information low-dimensional vector of content to be recommended Content construction DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

In some embodiments, it is special to calculate content in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Levy the vector similarity between information vector, comprising:

Calculate the vector length of the user's characteristic information vector sum content characteristic information vector；

Calculate the inner product of vectors between the user's characteristic information vector sum content characteristic information vector；

The COS distance of the content characteristic information vector is calculated based on the inner product of vectors and vector length；

The COS distance is normalized, user's characteristic information vector and content characteristic information vector are obtained Vector similarity.

The embodiment of the present application also provides a kind of content recommendation device, comprising:

Acquiring unit, it is described for obtaining the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes that the mapping between content to be recommended and the content characteristic information vector of content to be recommended is closed System；

Map unit obtains the user's characteristic information for carrying out feature space mapping to the user's characteristic information Corresponding user's characteristic information vector；

Computing unit, for calculating content characteristic in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Vector similarity between information vector；

Determination unit, for determining object content feature letter from content characteristic information vector based on the vector similarity Cease vector；

Selection unit, it is corresponding for choosing object content characteristic information vector based on the content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Object content；

Recommendation unit, for recommending the content of the object content to target user.

In some embodiments, the user's characteristic information includes user characteristics content and user characteristics vector, described to reflect Penetrating unit includes:

Subelement is mapped, for into default vector space, obtaining in user characteristics the user characteristics content map Hold vector；

Subelement is merged, for the user characteristics vector and user characteristics content vector to be carried out Vector Fusion, is obtained The corresponding user's characteristic information vector of the user's characteristic information.

In some embodiments, the mapping subelement is specifically used for:

The embodiment of the present application also provides a kind of storage medium, and the storage medium is stored with a plurality of instruction, and described instruction is suitable It is loaded in processor, to execute the step in any content recommendation method provided by the embodiment of the present application.

The embodiment of the present application also provides a kind of server, including memory is stored with a plurality of instruction；The processor is from institute It states and loads instruction in memory, to execute the step in any content recommendation method provided by the embodiment of the present application.

The user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of the available target user of the embodiment of the present application, Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes that the mapping between content to be recommended and the content characteristic information vector of content to be recommended is closed System；Feature space mapping is carried out to user's characteristic information, obtains the corresponding user's characteristic information vector of user's characteristic information；It calculates Vector similarity in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector；It is based on Vector similarity determines object content characteristic information vector from content characteristic information vector；Based on content DUAL PROBLEMS OF VECTOR MAPPING set of relations It closes and chooses the corresponding object content of object content characteristic information vector；Recommend object content to target user.In this application, may be used Using chosen by the user's characteristic information vector of target user content to be recommended similar with target user's characteristic information as Object content, to promote the accuracy of commending contents.

Detailed description of the invention

In order to more clearly explain the technical solutions in the embodiments of the present application, make required in being described below to embodiment Attached drawing is briefly described, it should be apparent that, the drawings in the following description are only some examples of the present application, for For those skilled in the art, without creative efforts, it can also be obtained according to these attached drawings other attached Figure.

Fig. 1 a is the schematic diagram of a scenario of content recommendation method provided by the embodiments of the present application；

Fig. 1 b is the first flow diagram of content recommendation method provided by the embodiments of the present application；

Fig. 1 c is the basic structural schematic diagram of the DSSM model of content recommendation method provided by the embodiments of the present application；

Fig. 1 d is the relationship signal between the included angle cosine value of content recommendation method provided by the embodiments of the present application and vector Figure；

Fig. 2 a is the system structure diagram of content recommendation method provided by the embodiments of the present application；

Fig. 2 b is second of flow diagram of content recommendation method provided by the embodiments of the present application；

Fig. 2 c is the model structure schematic diagram of the DSSM of content recommendation method provided by the embodiments of the present application；

Fig. 2 d is the network architecture schematic diagram of the DSSM of content recommendation method provided by the embodiments of the present application；

Fig. 3 a is the first structural schematic diagram of content recommendation device provided by the embodiments of the present application；

Fig. 3 b is second of structural schematic diagram of content recommendation device provided by the embodiments of the present application；

Fig. 4 is the structural schematic diagram of server provided by the embodiments of the present application.

Specific embodiment

Below in conjunction with the attached drawing in the embodiment of the present application, technical solutions in the embodiments of the present application carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of embodiments of the present application, instead of all the embodiments.It is based on Embodiment in the application, those skilled in the art's every other implementation obtained without creative efforts Example, shall fall in the protection scope of this application.

The embodiment of the present application provides a kind of content recommendation method, device, server and storage medium.

Wherein, which specifically can integrate in the electronic device, which can be terminal, service The equipment such as device.Wherein, terminal can be mobile phone, tablet computer, smart bluetooth equipment, laptop or PC Equipment such as (Personal Computer, PC).Server can be single server, be also possible to be made of multiple servers Server cluster.In one embodiment, which can also be integrated in multiple electronic equipments, for example, interior Holding recommendation apparatus can also be integrated in multiple servers, and present context recommended method is realized by multiple servers.

With reference to Fig. 1 a, which can be server, the user's characteristic information of the available user terminal of the server, And content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship include the content of content to be recommended Yu content to be recommended Mapping relations between characteristic information vector；Feature space mapping is carried out to user's characteristic information, obtains user's characteristic information pair The user's characteristic information vector answered；Calculate content characteristic information in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Vector similarity between vector；Based on vector similarity from content characteristic information vector determine object content characteristic information to Amount；The corresponding object content of object content characteristic information vector is chosen based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship；To target user Recommend object content.

It is described in detail separately below.It should be noted that the serial number of following embodiment is not as preferably suitable to embodiment The restriction of sequence.

In the present embodiment, a kind of content recommendation method is provided, which can be by server or terminal To execute.The embodiment of the present application will be introduced for being executed by server, specifically, can be held by the processor of server Row.As shown in Figure 1 b, the detailed process of the content recommendation method can be such that

101, the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, content DUAL PROBLEMS OF VECTOR MAPPING are obtained Set of relationship includes the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended.

Wherein, target user (Target customer), which refers to, provides the content of service or product using for it.For example, Target user can refer to currently in the upper user for registering, logging in and/or obtain service of goods of application.For example, target user can To be the user for being currently accessing application, currently log in the user, etc. of the application.

Wherein, user's characteristic information (Customer Information) is that target user carries personal letter in a network The data of breath.For example, classifying according to information type, user's characteristic information may include user description information, user behavior Information and subscriber association information.For example, user's characteristic information may include user's gender, historical viewings record, frequent contact, Etc..

For example, user description information can be the static data for reflecting user's essential attribute, such as gender, the surname of user Name, home address, income, personal label etc..User description information may come from the upload data of user itself, can also be with Upload data from other users, or carry out the collected user basic information of automatic network, etc..

For example, user behavior information can be the dynamic data for reflecting user behavior, preference, demand, for example, user purchases Buy record, spending amount, the historical viewings record etc. of service or product.Behavioral data can come from storage user behavior information Database, can be from the real-time record that local memory summarizes.

For example, subscriber association information refers to, reflection relevant to user behavior and influences the factors such as user behavior and psychology Related data.For example, user satisfaction, consumer loyalty degree, user's lifetime value etc..

Wherein, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, which refers to, contains the set of at least one mapping relations, the mapping relations Reciprocal correspondence, the relationship of projection of (Relational Mapping) between content to be recommended and content characteristic information vector.

In some embodiments, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship can be saved with various data structures, for example, according to number According to the difference of structure type, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship can be with chained list, table, hash table (hash), array, tree (tree) etc. forms save.

For example, reference table 1, table 1 provide a kind of content DUAL PROBLEMS OF VECTOR MAPPING set of relationship saved in a table format:

Number	Content to be recommended	Content characteristic information vector
			0x00	Content A	Vector a
0x01	Content B	Vector b
			0x02	Content C	Vector c

Table 1

The content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of above-mentioned table 1 contains 3 mapping relations set, and respectively number 0x00's reflects Penetrate the mapping relations [content of relationship [content A --- vector a], the mapping relations of 0x01 [content B --- vector b] and 0x02 C --- vector c].

Wherein, content to be recommended refers to that waiting is personalized the things or virtual things of recommendation.For example, content to be recommended can To refer to the virtual item data that can be recommended in database, and the related data of virtual things.For example, according to type point Class, content to be recommended may include event to be recommended (item) news such as to be recommended, hot list and article to be recommended (Object) commodity such as to be recommended, service, video, music.

Wherein, content characteristic information vector refers to the vector of characteristic information entrained by content to be recommended.For example, content characteristic Information vector can refer to it is being extracted from the information content entrained by content to be recommended, can express content characteristic to be recommended to Amount.

For example, content characteristic information vector can be in text content if content to be recommended carries content of text The term vector (Word embedding) of keyword, alternatively, if content to be recommended carries two dimensional image content, it is to be recommended interior The feature vector of appearance can be the textural characteristics, etc. that the two dimensional image is saved with vector format.

In some embodiments, the user's characteristic information for obtaining target user may include step in detail below:

A. user's request of target user is obtained；

B. inquiry user's characteristic information is requested according to user.

Wherein, user's request includes the identity information of request type and user, and request type is used to show the class of the request Type, subscriber identity information are the identity (identity, ID) for issuing the user of request, for example, subscriber identity information can be use The title at family, the number of user, network address of user etc..

For example, the format of user's request can be shown in reference table 2:

Request type	Customs Assigned Number	User's name
			0x01	20150412092236	hello_world_007

Table 2

Obtain user request mode have it is a variety of, for example, directly read from local memory storage user request, connect Receive user's request, etc. that user terminal is sent.For example, the Video Applications can take from trend when user logs in Video Applications Business device sends user's request.

Pass through the identity information of user in inquiry user's request, the user's characteristic information of the available user.Wherein, it looks into The method of inquiry have it is a variety of, for example, sending the identity information of the user to network database servers by network, then obtain The user's characteristic information that the database server returns；For another example, the subscriber identity information pair is directly read in local memory The user's characteristic information answered.

In some embodiments, content characteristic information includes the feature of content to be recommended, content feature vector, In, feature is the content characteristic information saved in the content characteristic information of content to be recommended with scalar likeness in form, and content is special Levying vector is the content characteristic information saved in the form of vectors；When the content characteristic for existing simultaneously vector form and non-vector form When information, for obtaining step content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, 101 can specifically include following sub-step:

(1) user's characteristic information of target user and the content characteristic information of content to be recommended are obtained.

Wherein, content characteristic information is to characteristic information entrained by sharing contents.For example, classify according to information type, Content characteristic information may include description information, the related information etc. to sharing contents.For example, content characteristic information may include Word content to sharing contents, the association to sharing contents share record to sharing contents, to the history of sharing contents, etc. Deng.

For example, when sharing contents are video content, the content characteristic information to sharing video frequency content may include this To the video-audio data of sharing video frequency content, video type, video tab, video name, video profile, viewing record, click volume, Uplink time, etc..

(2) feature is mapped in default vector space, obtains feature vector.

Wherein, default vector space refers to the linear space (vector space) of default dimension, for example, default vector is empty Between can for default dimension, can accommodate and support in the space of the movement or transformation that wherein occur.

For example, default vector space RⁿFeature is mapped to default vector by the real vector space for being n for a dimension Mode in space, which can be, is mapped to feature wherein by way of Linear Mapping.

For example, feature is mapped to by deep neural network (Deep Neural Network, DNN) preset to In quantity space, feature vector is obtained；Again for example, by feature operator (for example, local binary operator (Local Binary Pattern, LBP), Laplce (Laplace) operator, etc.) feature is mapped in default vector space, obtain spy Levy content vector, etc..

(3) the feature vector of content to be recommended and content feature vector are subjected to Vector Fusion, obtained in be recommended The content characteristic information vector of appearance.

Wherein, Vector Fusion refers to that multiple vectors are converted into a vector；According to the classification of amalgamation mode, Vector Fusion can be with Including vector splicing, addition of vectors, multiplication of vectors, etc..

For example, can using proper orthogonal decomposition by the feature vector of content to be recommended and content feature vector carry out to Amount splicing, to obtain the content characteristic information vector of content to be recommended.

In some embodiments, content characteristic information vector includes content characteristic information high dimension vector, in order to obtain content The feature vector of content feature vector and content to be recommended is carried out Vector Fusion by characteristic information low-dimensional vector, obtain to After the content characteristic information vector of recommendation, specifically can with the following steps are included:

Dimension-reduction treatment is carried out to content characteristic information high dimension vector, obtains content characteristic information low-dimensional vector.

Wherein, dimension-reduction treatment refers to using certain mapping method, by the Mapping of data points in former higher dimensional space to low dimensional Space in.The method of dimension-reduction treatment have it is a variety of, for example, according to the classification of type of dimension reduction method, dimension reduction method may include Explicit dimension reduction method, implicit dimension reduction method, linear dimension reduction method, Method of Nonlinear Dimensionality Reduction, etc..

And by dimension-reduction treatment, it may desire to error caused by the redundancy in content characteristic information high dimension vector, Retain the essential structure feature inside content characteristic information high dimension vector simultaneously, to improve the precision of commending contents.

There are many ways to dimension-reduction treatment, for example, the method for dimension-reduction treatment may include Principal Component Analysis Algorithm (Principal Component Analysis, PCA), linear dimension-reduction algorithm (Linear Discriminant Analysis, LDA), it is locally linear embedding into (Locally linear embedding, LLE), laplacian eigenmaps (Laplacian Eigenmaps), etc..

(4) it is constructed according to the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

For example, in some embodiments, can according to the content characteristic information of content to be recommended and content to be recommended to Mapping relations between amount carry out content construction DUAL PROBLEMS OF VECTOR MAPPING set of relationship with various data structures, for example, according to data structure class The difference of type can carry out content construction DUAL PROBLEMS OF VECTOR MAPPING set of relations in the form of chained list, table, hash table (hash), array, tree etc. It closes.

In some embodiments, content characteristic information vector includes content characteristic information high dimension vector, according to be recommended interior Mapping relations content construction DUAL PROBLEMS OF VECTOR MAPPING set of relationship between appearance and the content characteristic information vector of content to be recommended is specific It may comprise steps of:

According to the mapping relations building between content to be recommended and the content characteristic information low-dimensional vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

102, feature space mapping is carried out to user's characteristic information, obtains the corresponding user's characteristic information of user's characteristic information Vector.

Wherein, the mode for carrying out feature space mapping to user's characteristic information can refer to above-mentioned steps 101 (2) for feature Content map obtains the mode of feature vector into default vector space.

Specifically, in some embodiments, user's characteristic information includes user characteristics content and user characteristics vector, step 102 can specifically include following sub-step:

(1) by user characteristics content map into default vector space, user characteristics content vector is obtained.

It, specifically can be with for example, obtain user characteristics content vector by user characteristics content map into default vector space The following steps are included:

A. word segmentation processing is carried out to user characteristics content, obtains paragraph sequence, paragraph sequence includes at least a kind of paragraph.

Wherein, participle, which refers to, is divided into multiple character string subsequences according to certain specification for continuous character string sequence Process.The mode of participle have it is a variety of, for example, the segmenting method based on string matching, segmenting method and base based on understanding In the segmenting method, etc. of statistics.

For example, user characteristics content is " i have a coin ", it is based on character string dictionary, it can will be in the user characteristics Hold participle and obtains paragraph sequence [i] [have] [a] [coin].

B. the frequency of occurrences of the every kind of paragraph in paragraph sequence is counted, the corresponding paragraph frequency of every kind of paragraph is obtained.

In some embodiments, in order to solve to obtain, user characteristics content vector excessively coefficient, parameter space are excessive asks Topic, can count every kind of paragraph in paragraph sequence by N-gram (a kind of language model (Language Model, LM)) The frequency of occurrences, obtain the corresponding paragraph frequency of every kind of paragraph.

Wherein, N-gram may include 1-Gram (linear model), Bi-Gram (binary model), Tri-Gram (ternary mould Type), etc..

For N-Gram based in hidden Markov it is assumed that in i.e. one section of text, the probability that passage occurs is equal to should The probability product that each word occurs in section text.The probability of each word can be by being calculated in corpus.

Such as, it is assumed that passage T is by c₁, c₂, c₃...c_iComposition, it is as follows to be formulated 1-Gram language model:

Bi-Gram model is as follows:

Tri-Gram model is as follows:

Wherein, P (T) is the probability that passage occurs, and P (ci) is the probability that i-th of word occurs, and wherein p (X | Y) refers to The number that Y occurs in number/corpus that text X and Y occurs simultaneously in corpus.

Therefore first user characteristics content carries out word segmentation processing, obtains paragraph sequence, paragraph sequence includes at least a kind of paragraph, is adopted With the above-mentioned N-Gram formula, that is, frequency of occurrences of the statistics available every kind of paragraph in paragraph sequence, the corresponding paragraph of every kind of paragraph is obtained Frequency.

For example, after segmenting to this section of text, 4 paragraphs are obtained for passage [I have a coin], point Not Wei [I] [have] [a] [coin], using 1-Gram model calculate P (I, have, a, coin) known to:

P (I, have, a, coin)=P (I) P (have | I) P (a | have) P (coin | a)

C. the paragraph that paragraph frequency meets preset threshold is chosen from paragraph sequence, obtains target paragraph.

In order to further decrease the information redundancy in vector, paragraph frequency can be chosen from paragraph sequence and meet default threshold The paragraph of value, to filter the low paragraph of the frequency of occurrences.

For example, in some embodiments, the paragraph that paragraph frequency is greater than preset threshold being chosen from paragraph sequence, is obtained To target paragraph.

Wherein, which can be set as desired by technical staff.

D. the corresponding user's characteristic information vector of user's characteristic information is constructed according to target paragraph.

Finally, a frequency meter can be constructed according to the frequency of target paragraph, it is one-dimensional as eigenmatrix, then by the matrix Change, the corresponding user's characteristic information vector of user's characteristic information can be obtained.

(2) user characteristics vector and user characteristics content vector are subjected to Vector Fusion, it is corresponding obtains user's characteristic information User's characteristic information vector.

Wherein, the concrete mode of Vector Fusion can refer to step 101 (2) for content characteristic information vector and feature Vector carries out Vector Fusion, obtains the mode of the content characteristic information vector of content to be recommended, this will not be repeated here.

In some embodiments, user's characteristic information vector includes user's characteristic information high dimension vector, and step 102 specifically may be used To include following sub-step:

Dimension-reduction treatment is carried out to user's characteristic information high dimension vector, obtains user's characteristic information low-dimensional vector.

Wherein, the method for above-mentioned dimension-reduction treatment specifically includes:

(1) connection weight and biasing in Matching Model is used to be weighted summation to user's characteristic information high dimension vector Processing, user's characteristic information high dimension vector after being handled, Matching Model is by being labelled with the training sample of true vector similarity Training forms.

Wherein, which can be semantic model (the Deep Structured Semantic based on depth network Model, DSSM), which can be by the language of user's characteristic information and content characteristic information MAP a to identical dimensional Adopted space finally calculates the cosine between the two vectors to obtain user's characteristic information vector sum content characteristic information vector (cosine) distance obtains vector similarity.

Specifically, DSSM model may include common DSSM model, can also include the DSSM model based on convolution (Convolutional DSSM, C-DSSM), circulation DSSM model (Recurrent DSSM, R-DSSM), multiple information sources DSSM Model (Multi-View DSSM, MV-DSSM), the multi-modal scale model of deep layer (Deep Multimodal Similarity Model, DMSM), etc..

Wherein, the basic structure of DSSM model can refer to Fig. 1 c, and as illustrated in figure 1 c, DSSM model may include matching Layer, expression layer and input layer.

Wherein, input layer can in expression layer, then expression layer for inputting user's characteristic information A and content characteristic information B User's characteristic information A is projected as user's characteristic information vector a, by content characteristic information B be projected as content characteristic information to Measure b, then, can be calculated in matching layer the cosine between user's characteristic information vector a and content characteristic information vector b away from From, will the COS distance normalization after obtain vector similarity.

For example, in some embodiments, using the connection weight in Matching Model and biasing to user's characteristic information height Dimensional vector is weighted summation process, after being handled before user's characteristic information high dimension vector, can also specifically include following Step:

A. training sample is obtained, training sample includes user's training sample, content training sample, and training sample is labelled with use True vector similarity between family training sample and content training sample；

B. initial matching model is trained using user's training sample, content training sample, obtains user's training sample Predicted vector similarity between sheet and content training sample；

C. initial matching model is restrained according to true vector similarity and predicted vector similarity, after being trained Matching Model.

Obtain training sample mode have it is a variety of, for example, obtaining trained sample from web data server by network Originally, training sample and the acquisition training sample, etc. in local memory are directly read.The training sample can be by art personnel Mark the true vector similarity between user's training sample and content training sample.

In some embodiments, in order to guarantee the vector similarity acquired closer to 1, user's characteristic information and to be recommended interior The content characteristic information of appearance is more similar, using the connection weight in Matching Model and biases to user's characteristic information high dimension vector It is weighted summation process, user's characteristic information high dimension vector may include specific steps below after being handled:

A. when the connection weight in Matching Model is nonnegative number, according to connection weight and biasing to user's characteristic information High dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled；

B. when the connection weight in Matching Model is negative, according to the absolute value of connection weight and biasing to user spy Reference breath high dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled.

In some embodiments, the absolute value to the non-negative limitation of connection weight in addition to obtaining connection weight, can also lead to Numerical intervals where crossing judgement connection weight come to the connection weight assignment.

For example, connection weight is assigned a value of 15 when connection weight numerical value is in [- 20,10] section.

Specific assignment mode can be arranged by technical staff.

(2) user's characteristic information high dimension vector after processing is carried out at dimensionality reduction using the default dimensionality reduction function in Matching Model Reason, obtains user's characteristic information low-dimensional vector.

Wherein, presetting dimensionality reduction function is the dimensionality reduction function in Matching Model in presentation layer, for example, when Matching Model is DSSM When model, which can be principal component analysis (Principal Component Analysis, PCA) function.

For example, setting n-dimensional vector w as a change in coordinate axis direction (referred to as map vector) of target subspace, data are maximized Variance after mapping, has:

Wherein m is the number of data instance, and value can be positive integer, x_iIt is the vector expression of data instance i, i's Value can be positive integer,It is the average vector of all data instances.W is defined to be column vector comprising all map vectors Matrix, by linear algebraic transformation, available following optimization object function:

The wherein mark number (trace) of tr representing matrix, A is data covariance matrix.

Known to ground, optimal W is by the corresponding feature vector of the maximum characteristic value of k before data covariance matrix as column vector It constitutes, this feature vector forms one group of orthogonal basis and retains vector information.

The output of PCA is Y=W`X, wherein is reduced to k dimension by the original dimension of X.

103, it calculates in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector Vector similarity.

Wherein, vector similarity (Similarity Measurement) can be by calculating the distance between vector (distance) it acquires.

For example, Euclidean distance (Euclidean Distance), manhatton distance (Manhattan can be passed through Distance), Chebyshev's distance (Chebyshev Distance), standardization Euclidean distance (Standardized Euclidean distance), included angle cosine (Cosine) etc. calculate vector similarity.

In some embodiments, content characteristic information vector includes content characteristic information low-dimensional vector, and step 103 specifically may be used With the following steps are included:

It calculates in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information low-dimensional vector Vector similarity.

For example, by normalization included angle cosine value as vector similarity:

Wherein, with reference to the relation schematic diagram between Fig. 1 d included angle cosine value provided and vector, a, b are two vectors, cos θ is included angle cosine value (being also COS distance, cosine similarity), and i and n are positive integer, and the calculation of vector similarity is The included angle cosine value is normalized by softmax function, obtains the numerical value between [0,1], wherein vector similarity is such as Under:

104, object content characteristic information vector is determined from content characteristic information vector based on vector similarity.

For example, in some embodiments, determining that object content is special from content characteristic information vector based on vector similarity The mode for levying information vector is sorted from large to small to vector similarity, determines preceding 100 in content characteristic information vector It is a to be used as object content characteristic information vector.

For example, in some embodiments, determining that object content is special from content characteristic information vector based on vector similarity Sign information vector mode be obtain all the elements characteristic information vector in vector similarity greater than 0.8 content characteristic information to Amount, as object content characteristic information vector.

For example, in some embodiments, determining that object content is special from content characteristic information vector based on vector similarity The mode for levying information vector is sorted from large to small to vector similarity, determines preceding 100 in content characteristic information vector A middle vector similarity is greater than 0.8 content characteristic information vector, as object content characteristic information vector.

105, the corresponding object content of object content characteristic information vector is chosen based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

For example, choosing should according to the mapping relations of object content characteristic information vector in content DUAL PROBLEMS OF VECTOR MAPPING set of relationship The corresponding object content of object content characteristic information vector.

106, recommend the content of object content to target user.

For example, the object content to be transmitted directly to the user terminal of target user.

For example, in some embodiments, can also further be segmented, be filtered to the object content, then by its name Claim the user terminal for being sent to target user.

For example, in some embodiments, the title of the object content can be searched in database server by network, The other information of the object content is obtained, the object content and other information are sent jointly to the user terminal of target user.

For example, the mesh can be searched in database server by network when object content is target video content The title for marking video content, obtains the other information such as click volume, comment, temperature, the brief introduction of the target video content, by the target Content and other information send jointly to the user terminal of target user.

Commending contents scheme provided by the embodiments of the present application can be applied in various commending contents scenes, such as.With view For frequency is launched, server can log in the user of acquisition target user when video playing is applied in target user by user terminal Characteristic information (for example, user's gender, history video tour record, age of user, user tag, etc.), and read local Video content DUAL PROBLEMS OF VECTOR MAPPING set of relationship in memory.Wherein, video content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes video to be recommended With the mapping relations between video content features information vector to be recommended.

Then, server can carry out feature space mapping to user's characteristic information, and it is corresponding to obtain user's characteristic information User's characteristic information vector, and calculate video to be recommended in user's characteristic information vector and video content DUAL PROBLEMS OF VECTOR MAPPING set of relationship The vector similarity of content characteristic information vector.

Determine target video content characteristic letter from video content features information vector to be recommended based on vector similarity again Vector is ceased, and the corresponding recommendation view of target video content characteristic information vector is chosen based on video content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Frequently, video playing application homepage finally is logged in the user terminal of target user launch the recommendation video.

For example, the characteristic information of user is [women, 23 years old, historical viewings makeups video, user tag: lipstick tries color], By carrying out feature space mapping to the user's characteristic information, user's characteristic information vector is obtained, and according to the user characteristics Information vector retrieves makeups video content features information vector similar with its from local memory, then to the view of the user terminal Frequency, which is played, launches the makeups video retrieved using homepage.

Using scheme the embodiment of the present application provided by the embodiments of the present application by obtain feature on user's characteristic information phase As content to be recommended as object content, the accuracy of commending contents can be promoted.

From the foregoing, it will be observed that the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING of the available target user of the embodiment of the present application Set of relationship, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship include between content to be recommended and the content characteristic information vector of content to be recommended Mapping relations；Feature space mapping is carried out to user's characteristic information, obtains the corresponding user's characteristic information of user's characteristic information Vector；Calculate the vector phase between user's characteristic information vector and content characteristic information vector in content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Like degree；Object content characteristic information vector is determined from content characteristic information vector based on vector similarity；Based on content vector Mapping relations set chooses the corresponding object content of object content characteristic information vector；Recommend object content to target user.By This, the embodiment of the present application, can be with by obtaining in feature content to be recommended similar with user's characteristic information as object content Promote the accuracy of commending contents.

The method according to described in above-described embodiment, will now be described in further detail below.

In the present embodiment, it will be integrated in content recommendation device in offline service device and line server, and carry out video For commending contents, the method for the embodiment of the present invention is described in detail.Offline service device and online in one embodiment Server can also be realized by a server.

Fig. 2 a is a kind of system structure diagram of commending contents provided in this embodiment, as shown in Figure 2 a, the present embodiment Content recommendation device can be integrated in respectively offline service device from different information are handled in line server.

Wherein, it can store the attribute information of video in the database of offline service device.Wherein, the attribute information of video can User, video content types, etc. are watched with the address information, title, label, the history that include video.

In offline service device, training data is can be used to train initial Matching Model in content recommendation device, and fixed Matching Model after training is sent to line server by the phase.

Then, it is mentioned by Matching Model after training to carry out feature to the attribute information that can store video in database It takes, obtains the feature vector of video to be recommended, and the feature vector of video to be recommended is periodically sent to line server.

In line server, content recommendation device can periodically obtain after the training of update Matching Model and to be recommended Content characteristic information vector, and Matching Model mentions after the training updated by these to carry out feature to user's characteristic information It takes, obtains user's characteristic information vector, then retrieval and user's characteristic information in all content characteristic information vectors to be recommended The corresponding content to be recommended of vector most like first 100 content characteristic information vectors to be recommended, as object content.

As shown in Figure 2 b, a kind of content recommendation method can be executed, detailed process by offline service device and line server It is as follows:

201, offline service device training initial matching model, Matching Model after being trained.

In offline service device, the available training data of content recommendation device, the training data includes user's training sample Originally, content training sample, the training data are also labelled with the true vector phase between user's training sample and content training sample Like degree.

Initial matching model can be trained according to these training data offline service devices, and mould will be matched after training Type is periodically uploaded to line server.

In the present embodiment, content recommendation device can obtain training number by network from network database servers According to.

Content recommendation device can read initial matching model in local memory, which is depth characteristic Matching Model (Deep Semantic Matching Model, DSSM), the model structure schematic diagram of the model can be with reference to figure 2c, as shown in Figure 2 c, DSSM include three network layers, respectively input layer, expression layer and matching layer.

The training data that input layer can will acquire imports DSSM.

Expression layer can will extract the high dimensional feature vector of training data, and respectively by these high dimensional feature vectors by more Layer perceptron (Multi-Layer Perception, MLP) Lai Jinhang dimensionality reduction, obtains low-dimensional feature vector.

The cosine value that matching layer can carry out the low-dimensional vector of data between vector two-by-two calculates, and the cosine value acquired is led to It crosses softmax function to be normalized, obtains predicted vector similarity.

Finally, according to the true vector similarity and predicted vector similarity marked in training data to initial matching model It is restrained, Matching Model after being trained.

Matching Model not only can predict the vector similarity between two word contents in matching layer after the training, but also can be with In the corresponding low-dimensional feature vector of expression layer output character content.

Specifically, with reference to the network architecture schematic diagram of Fig. 2 d DSSM provided a kind of, as shown in Figure 2 d, DSSM every time can be with One user's training sample of input layer input and multiple content training samples are converted to the high dimensional feature vector of 500k dimension, Obtain the user training high dimensional feature vector Q and content training high dimensional feature vector D of one-hot encoding (one-hot) form₁、 D₂...D_n, specific transform mode can be with reference to step 102, and this will not be repeated here.

Since vector dimension is excessively high, calculation amount is excessive, thereby increases and it is possible to it will lead to OOV (Out of Vocabulary) problem, therefore Word Hash (Word Hashing) operation is also needed to be implemented in expression layer, to reduce the dimension of high dimensional feature vector, generates 30k dimension The feature vector of degree.

Then, the feature vector of 30k is generated into the low-dimensional feature vector of 128 dimensions by MLP.

Finally, the content in matching layer by user's training low-dimensional feature vector Q of 128 dimensions respectively with 128 dimensions trains low-dimensional Feature vector D₁、D₂...D_nThe calculating of vector cosine similarity is carried out, D is obtained₁Cosine similarity R (Q, D₁)、D₂Cosine it is similar Spend R (Q, D₂)...D_nCosine similarity R (Q, D_n)。

Also, these cosine similarities are normalized by softmax function, are converted it into big less than 1 In 0 probability value, D is obtained₁Vector similarity P (Q | D₁)、D₂Vector similarity P (Q | D₂)...D_nVector similarity P (Q |D_n), vector similarity is closer to 1, then user's training sample is more similar to content training sample.

Word hashing operation can refer to user characteristics content map in step 102 into vector space, obtain user spy The step of levying content vector, this will not be repeated here.

Wherein, the specific structure of multilayer perceptron can carry out user's characteristic information high dimension vector with reference in step 102 Dimension-reduction treatment, the step of obtaining user's characteristic information low-dimensional vector, this will not be repeated here.

Finally, minimizing loss function using Maximum-likelihood estimation in training stage the application:

Residual error can the backpropagation in the deep neural network of expression layer, eventually by stochastic gradient descent (Stochastic Gradient Descent, SGD) restrains model, obtains the parameter { Wi, bi } of each network layer.

202, offline service device obtains the content characteristic information of content to be recommended, and using Matching Model after training by content Feature information processing is constructed at content characteristic information low-dimensional vector, and according to the content characteristic information low-dimensional vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

In the present embodiment, offline service device can also be obtained by network from network database servers in be recommended The content characteristic information of appearance, and believed according to Matching Model after training in step 201 come the content characteristic to these contents to be recommended Breath processing, generates content low-dimensional language feature vector.

Wherein, content information to be recommended includes the information such as number, title, address, the label of content to be recommended.

Then, the corresponding content low-dimensional language feature vector of these contents to be recommended is constructed into content DUAL PROBLEMS OF VECTOR MAPPING set of relations It closes, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship format shown in reference table 3, which may include multiple Content to be recommended and the corresponding content low-dimensional language feature vector of each content to be recommended, url (Uniform Resource Locator) address and title.

Content to be recommended	Title	Content low-dimensional language feature vector	The address url
				Content A to be recommended	Landscape painting a	A [x, x, x]	http://xxx
Content B to be recommended	Landscape video b	B [x, x, y]	http://xxy
				Content C to be recommended	Travel video c	C [x, y, y]	http://xyy

Table 3

203, line server obtains user's characteristic information.

In the present embodiment, the available recommendation request from user terminal of line server, which includes user Identity information, address information etc..

Inquire the identity information of user in network database servers by network, available user's characteristic information, The user's characteristic information includes the information such as gender, age, the historical viewings record of user.

204, line server carries out feature space mapping to user's characteristic information using model after training, obtains user spy Reference ceases corresponding user's characteristic information high dimension vector, and carries out dimension-reduction treatment to user's characteristic information high dimension vector, is used Family characteristic information low-dimensional vector.

In line server, content recommendation device matches after can obtaining the training that offline service device is sent by network Model, and feature space mapping is carried out to user's characteristic information using the input layer of model after training and presentation layer, obtain table The user's characteristic information high dimension vector exported up to layer.

Wherein, for the word content in user's characteristic information, the presentation layer of Matching Model can be executed following after training Word hashing step, to be translated into vector, and solve the problems, such as data scale, improve robustness:

A. slicing treatment is carried out to user characteristics content, obtains paragraph sequence；

B. the frequency of occurrences for counting every kind of paragraph in paragraph sequence, obtains the corresponding paragraph frequency of every kind of paragraph；

C. the paragraph that paragraph frequency meets preset threshold is chosen from paragraph sequence, obtains target paragraph；

D. the corresponding feature vector of word content is constructed according to target paragraph.

The one-hot encoding vector of higher-dimension can be converted to how hot code (multi-hot) by the present embodiment as a result, for example, by literary Word content using Tri-gram be sliced to obtain the word content for paragraph sequence, so that it is corresponding special to compress word content Levy the space of vector.

For example, word content #Query#, can be sliced and generate a paragraph sequence, it include 5 in the paragraph sequence to language Section, respectively #Qu, Que, #uer, #ery, ry#.Then, these paragraphs are counted, every primary paragraph occurred and exist Its corresponding position+1, eventually forms the coding of a multi-hot.

In the present embodiment, these obtained feature vectors can also be stitched together, then, directly by user characteristics Feature vector in information is spliced after these feature vectors, and the corresponding user's characteristic information higher-dimension of user's characteristic information is obtained Vector.

Specific steps refer to step 102, and this will not be repeated here.

Again by the MLP network in presentation layer, user's characteristic information low-dimensional vector is obtained.With reference to Fig. 2 d, MLP mono- shares three Layer, W_iIndicate i-th layer of weight matrix, b_iIndicate i-th layer of biasing (bias), y is the vector that presentation layer finally exports, can be with It respectively indicates are as follows:

l₁=W₁x

l_i=f (W_il_i-1+b_i), i=2,3...N-1

Y=f (W_Nl_N-1+b_N)

Wherein, f is activation primitive, and in the present embodiment, f is tanh function:

205, it is special to calculate content in user's characteristic information low-dimensional vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship for line server Reference ceases the vector similarity between low-dimensional vector.

In line server, content recommendation device can obtain the content vector that offline service device is sent by network and reflect Set of relationship is penetrated, and using the matching layer of model after training come to user's characteristic information low-dimensional vector and content DUAL PROBLEMS OF VECTOR MAPPING All the elements characteristic information low-dimensional vector carries out length computation in set of relationship, inner product calculates, and obtains user's characteristic information low-dimensional The vector length of vector, the vector length of content characteristic information low-dimensional vector, user's characteristic information low-dimensional vector sum content characteristic Inner product of vectors between information low-dimensional vector.

Then, the vector length of vector length, content characteristic information low-dimensional vector based on user's characteristic information low-dimensional vector Inner product of vectors between degree, user's characteristic information low-dimensional vector sum content characteristic information low-dimensional vector, to calculate content characteristic letter Cease the COS distance of vector.

Finally, converting the COS distance to by softmax function the vector similarity of Probability Forms expression.

For example, the vector length of user's characteristic information low-dimensional vector Q is | Q |, the vector of content characteristic information low-dimensional vector D Length is | D |, the inner product of vectors between user's characteristic information low-dimensional vector Q and content characteristic information low-dimensional vector D is QD, COS distance are as follows:

Its vector similarity are as follows:

Wherein, the smoothing factor that wherein r is softmax, D are the positive sample under user's characteristic information low-dimensional vector, and D` is used Negative sample (take random negative sampling) under the characteristic information low-dimensional vector of family, all samples under D user's characteristic information low-dimensional vector This.

206, line server determines object content characteristic information based on vector similarity from content characteristic information vector Vector, and the corresponding object content of object content characteristic information vector is chosen, recommend object content to target user.

Finally, being ranked up the vector similarity being calculated is descending, chooses vector similarity and be greater than 0.9 and arrange Preceding 100 vector similarities of name, it is corresponding to be recommended in content DUAL PROBLEMS OF VECTOR MAPPING set of relationship to retrieve these vector similarities Title, the address url of content etc., as object content and the attribute of object content, then, by these object contents and mesh The attribute of mark content is sent to target user.

It in this application, can be by a portion target after executing fine sort, filtering sensitive words to these object contents The attribute of content and object content is sent to user terminal.

From the foregoing, it will be observed that the embodiment of the present application can train initial matching model, Matching Model after being trained is obtained wait push away Recommend the content characteristic information of content, and using Matching Model after training by content characteristic information processing at content characteristic information low-dimensional Vector, and according to the content characteristic information low-dimensional vector content construction DUAL PROBLEMS OF VECTOR MAPPING set of relationship of content to be recommended, obtain user Characteristic information carries out feature space mapping to user's characteristic information, and it is high to obtain the corresponding user's characteristic information of user's characteristic information Dimensional vector, and dimension-reduction treatment is carried out to user's characteristic information high dimension vector, user's characteristic information low-dimensional vector is obtained, user is calculated Vector similarity in characteristic information low-dimensional vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information low-dimensional vector, Object content characteristic information low-dimensional vector is determined from content characteristic information low-dimensional vector based on vector similarity, and chooses target The corresponding object content of content characteristic information low-dimensional vector recommends the object content to target user.

Due to the program be continuously updated training after Matching Model carry out and content characteristic information low-dimensional vector, can be improved The acquisition range of object content and the richness of object content, to meet the recommended requirements of different type user；Meanwhile passing through Calculate the vector similarity between low-dimensional feature vector, it is possible to reduce the calculation amount of similarity calculation, and line server is not necessarily to It is vector format by content transformation to be recommended, therefore calculation amount needed for can reducing retrieval in retrieval is to improve retrieval rate； Also, in the present embodiment, user's characteristic information vector considers more user's characteristic informations, thus, it is possible in being promoted Hold the accuracy recommended.

In order to better implement above method, the embodiment of the present application also provides a kind of content recommendation device, the commending contents Device specifically can integrate in the electronic device, which can be the equipment such as terminal, server, PC.For example, In the present embodiment, the method for the embodiment of the present invention will be carried out detailed so that content recommendation device is integrated in the server as an example Explanation.

For example, as shown in Figure 3a, which may include acquiring unit 301, map unit 302, calculates list Member 303, determination unit 304, selection unit 305 and recommendation unit 306 are as follows:

(1) acquiring unit 301:

Acquiring unit 301, for obtaining the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes the mapping relations between content to be recommended and content characteristic information vector.

In some embodiments, acquiring unit 301 may include obtaining subelement, content map subelement, content mergence Subelement and building subelement, as follows:

Subelement is obtained, for obtaining the user's characteristic information of target user and the content characteristic letter of content to be recommended Breath, content characteristic information includes feature, content characteristic information vector；

Content map subelement obtains feature vector for feature to be mapped in default vector space；

Content mergence subelement is obtained for content characteristic information vector and feature vector to be carried out Vector Fusion The content characteristic information vector of content to be recommended；

Subelement is constructed, for according between content to be recommended and the content characteristic information vector of content to be recommended Mapping relations content construction DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

In some embodiments, content characteristic information vector includes content characteristic information high dimension vector, and content mergence is single Member can be also used for:

Dimension-reduction treatment is carried out to content characteristic information high dimension vector, obtains content characteristic information low-dimensional vector；

According to the mapping relations content construction between content to be recommended and the content characteristic information vector of content to be recommended DUAL PROBLEMS OF VECTOR MAPPING set of relationship, comprising:

(2) map unit 302:

It is corresponding to obtain user's characteristic information for carrying out feature space mapping to user's characteristic information for map unit 302 User's characteristic information vector.

In some embodiments, user's characteristic information vector includes user's characteristic information high dimension vector, and map unit 302 can To be used for:

In some embodiments, user's characteristic information vector includes user's characteristic information high dimension vector, and map unit 302 can It is as follows to include weighting subelement and dimensionality reduction subelement:

(1) subelement is weighted, for using the connection weight in Matching Model and biasing to user's characteristic information higher-dimension Vector is weighted summation process, and user's characteristic information high dimension vector after being handled, Matching Model is by being labelled with true vector The training sample training of similarity forms.

In some embodiments, weighting subelement can be also used for:

Training sample is obtained, training sample includes user's training sample, content training sample, and training sample is labelled with user True vector similarity between training sample and content training sample；

Initial matching model is trained using user's training sample, content training sample, obtains user's training sample Predicted vector similarity between content training sample；

Initial matching model is restrained according to true vector similarity and predicted vector similarity, after being trained Matching Model.

(2) dimensionality reduction subelement, for high to user's characteristic information after processing using the default dimensionality reduction function in Matching Model Dimensional vector carries out dimension-reduction treatment, obtains user's characteristic information low-dimensional vector.

In some embodiments, dimensionality reduction subelement specifically can be used for:

In some embodiments, with reference to Fig. 3 b, map unit 302 may include mapping subelement 3021 and fusion subelement 3022, as follows:

(1) subelement 3021 is mapped:

Subelement 3021 is mapped, for into default vector space, obtaining in user characteristics user characteristics content map Hold vector.

In some embodiments, mapping subelement 3021 can be specifically used for:

Word segmentation processing is carried out to user characteristics content, obtains paragraph sequence, paragraph sequence includes at least a kind of paragraph；

The frequency of occurrences of the every kind of paragraph in paragraph sequence is counted, the corresponding paragraph frequency of every kind of paragraph is obtained；

The paragraph that paragraph frequency meets preset threshold is chosen from paragraph sequence, obtains target paragraph；

The corresponding user's characteristic information vector of user's characteristic information is constructed according to target paragraph.

(2) subelement 3022 is merged:

Subelement 3022 is merged, for user characteristics vector and user characteristics content vector to be carried out Vector Fusion, is obtained The corresponding user's characteristic information vector of user's characteristic information.

(3) computing unit 303:

Computing unit 303, for calculating content characteristic in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Vector similarity between information vector.

In some embodiments, computing unit 303 may include length subelement, inner product subelement, apart from subelement and Similarity subelement:

Length subelement, for calculating the vector length of user's characteristic information vector sum content characteristic information vector；

Inner product subelement, for calculating the inner product of vectors between user's characteristic information vector sum content characteristic information vector；

Apart from subelement, for calculating the COS distance of content characteristic information vector based on inner product of vectors and vector length；

Similarity subelement obtains user's characteristic information vector and content for COS distance to be normalized Vector similarity between characteristic information vector.

In some embodiments, content characteristic information vector includes content characteristic information low-dimensional vector, and computing unit 303 has Body can be used for:

(4) determination unit 304:

Determination unit 304, for determining object content feature letter from content characteristic information vector based on vector similarity Cease vector.

(5) selection unit 305:

Selection unit 305, it is corresponding for choosing object content characteristic information vector based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Object content.

(6) recommendation unit 306:

Recommendation unit 306, for recommending the content of object content to target user.

When it is implemented, above each unit can be used as independent entity to realize, any combination can also be carried out, is made It is realized for same or several entities, the specific implementation of above each unit can be found in the embodiment of the method for front, herein not It repeats again.

From the foregoing, it will be observed that the content recommendation device of the present embodiment is obtained the user's characteristic information of target user by acquiring unit, And content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship include the content of content to be recommended Yu content to be recommended Mapping relations between characteristic information vector；Feature space mapping is carried out to user's characteristic information by map unit, obtains user The corresponding user's characteristic information vector of characteristic information；User's characteristic information vector is calculated by computing unit and content DUAL PROBLEMS OF VECTOR MAPPING is closed Vector similarity in assembly conjunction between content characteristic information vector；Vector similarity is based on by determination unit to believe from content characteristic It ceases and determines object content characteristic information vector in vector；It is single based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship selection object content by choosing The corresponding object content of characteristic information vector；Recommend the content of object content from recommendation unit to target user.

Since the program can be chosen and target user's characteristic information by the user's characteristic information vector of target user Similar content to be recommended is as object content, and the present embodiment is so as to promoting the accuracies of commending contents.

The embodiment of the present application also provides a kind of server, which can be mobile phone, tablet computer, miniature handle box Son, unmanned plane or image capture device etc..As shown in figure 4, it illustrates servers involved in the embodiment of the present application Structural schematic diagram, specifically:

The server may include one or processor 401, one or more meters of more than one processing core The components such as memory 402, power supply 403, input module 404 and the communication module 405 of calculation machine readable storage medium storing program for executing.This field skill Art personnel are appreciated that server architecture shown in Fig. 4 does not constitute the restriction to server, may include more than illustrating Or less component, perhaps combine certain components or different component layouts.Wherein:

Processor 401 is the control centre of the server, utilizes each of various interfaces and the entire server of connection Part by running or execute the software program and/or module that are stored in memory 402, and calls and is stored in memory Data in 402, the various functions and processing data of execute server, to carry out integral monitoring to server.In some realities It applies in example, processor 401 may include one or more processing cores；In some embodiments, processor 401 can integrate at Manage device and modem processor, wherein the main processing operation system of application processor, user interface and application program etc. are adjusted Demodulation processor processed mainly handles wireless communication.It is understood that above-mentioned modem processor can not also integrate everywhere It manages in device 401.

Memory 402 can be used for storing software program and module, and processor 401 is stored in memory 402 by operation Software program and module, thereby executing various function application and data processing.Memory 402 can mainly include storage journey Sequence area and storage data area, wherein storing program area can the (ratio of application program needed for storage program area, at least one function Such as sound-playing function, image player function) etc.；Storage data area, which can be stored, uses created data according to server Deng.In addition, memory 402 may include high-speed random access memory, it can also include nonvolatile memory, for example, at least One disk memory, flush memory device or other volatile solid-state parts.Correspondingly, memory 402 can also include Memory Controller, to provide access of the processor 401 to memory 402.

Server further includes the power supply 403 powered to all parts, and in some embodiments, power supply 403 can pass through electricity Management system and processor 401 are logically contiguous, to realize management charging, electric discharge and power consumption by power-supply management system The functions such as management.Power supply 403 can also include one or more direct current or AC power source, recharging system, power supply event Hinder the random components such as detection circuit, power adapter or inverter, power supply status indicator.

The server may also include input module 404, which can be used for receiving the number or character letter of input Breath, and generation keyboard related with user setting and function control, mouse, operating stick, optics or trackball signal are defeated Enter.

The server may also include communication module 405, and in some embodiments, communication module 405 may include wireless son Module, server can carry out short range wireless transmission by the wireless submodule of the communication module 405, to provide wirelessly Broadband internet access.For example, the communication module 405 can be used for that user is helped to send and receive e-mail, browse webpage and access Streaming video etc..

Although being not shown, server can also be including display unit etc., and details are not described herein.Specifically in the present embodiment, Processor 401 in server can according to following instruction, by the process of one or more application program is corresponding can It executes file to be loaded into memory 402, and runs the application program being stored in memory 402 by processor 401, thus Realize various functions, as follows:

Obtain the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, content DUAL PROBLEMS OF VECTOR MAPPING relationship Set includes the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended；

To user's characteristic information carry out feature space mapping, obtain the corresponding user's characteristic information of user's characteristic information to Amount；

Calculate in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector to Measure similarity；

Object content characteristic information vector is determined from content characteristic information vector based on vector similarity；

The corresponding object content of object content characteristic information vector is chosen based on content DUAL PROBLEMS OF VECTOR MAPPING set of relationship；

Recommend object content to target user.

The specific implementation of above each operation can be found in the embodiment of front, and details are not described herein.

From the foregoing, it will be observed that the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING relationship of the available target user of the present embodiment Set, content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes reflecting between content to be recommended and the content characteristic information vector of content to be recommended Penetrate relationship；Feature space mapping is carried out to user's characteristic information, obtains the corresponding user's characteristic information vector of user's characteristic information； Calculate the vector similarity in user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector； Object content characteristic information vector is determined from content characteristic information vector based on vector similarity；It is closed based on content DUAL PROBLEMS OF VECTOR MAPPING Assembly, which is closed, chooses the corresponding object content of object content characteristic information vector；Recommend object content to target user.As a result, at this In application, it can be chosen by the user's characteristic information vector of target user similar with target user's characteristic information to be recommended Content is as object content, to promote the accuracy of commending contents.

It will appreciated by the skilled person that all or part of the steps in the various methods of above-described embodiment can be with It is completed by instructing, or relevant hardware is controlled to complete by instruction, which can store in computer-readable storage In medium, and is loaded and executed by processor.

For this purpose, the embodiment of the present application provides a kind of storage medium, wherein being stored with a plurality of instruction, which can be processed Device is loaded, to execute the step in any content recommendation method provided by the embodiment of the present application.For example, the instruction can To execute following steps:

Recommend object content to target user.

Wherein, which may include: read-only memory (ROM, Read Only Memory), random access memory Body (RAM, Random Access Memory), disk or CD etc..

By the instruction stored in the storage medium, any content provided by the embodiment of the present application can be executed and pushed away The step in method is recommended, it is thereby achieved that achieved by any content recommendation method provided by the embodiment of the present application Beneficial effect is detailed in the embodiment of front, and details are not described herein.

Above to a kind of content recommendation method, device, server and storage medium provided by the embodiment of the present application into It has gone and has been discussed in detail, specific examples are used herein to illustrate the principle and implementation manner of the present application, the above implementation The explanation of example is merely used to help understand the present processes and its core concept；Meanwhile for those skilled in the art, according to According to the thought of the application, there will be changes in the specific implementation manner and application range, and to sum up, the content of the present specification is not answered It is interpreted as the limitation to the application.

Claims

1. a kind of content recommendation method characterized by comprising

Obtain the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, the content DUAL PROBLEMS OF VECTOR MAPPING relationship Set includes the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended；

Feature space mapping is carried out to the user's characteristic information, obtains the corresponding user's characteristic information of the user's characteristic information Vector；

Calculate in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector to Measure similarity；

Recommend the object content to target user.

2. content recommendation method as described in claim 1, which is characterized in that the user's characteristic information includes in user characteristics Hold and it is corresponding to obtain the user's characteristic information to user's characteristic information progress feature space mapping for user characteristics vector User's characteristic information vector, comprising:

The user characteristics vector and user characteristics content vector are subjected to Vector Fusion, it is corresponding to obtain the user's characteristic information User's characteristic information vector.

3. content recommendation method as claimed in claim 2, which is characterized in that by the user characteristics content map to preset to In quantity space, user characteristics content vector is obtained, comprising:

4. content recommendation method as described in claim 1, which is characterized in that the content characteristic information vector includes content spy Reference ceases low-dimensional vector, calculate in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship content characteristic information to Vector similarity between amount, comprising:

It calculates in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information low-dimensional vector Vector similarity.

5. content recommendation method as described in claim 1, which is characterized in that the user's characteristic information vector includes user spy Reference ceases high dimension vector, carries out feature space mapping to the user's characteristic information, it is corresponding to obtain the user's characteristic information User's characteristic information vector, comprising:

Dimension-reduction treatment is carried out to the user's characteristic information high dimension vector, obtains user's characteristic information low-dimensional vector.

6. content recommendation method as claimed in claim 5, which is characterized in that carried out to the user's characteristic information high dimension vector Dimension-reduction treatment obtains user's characteristic information low-dimensional vector, comprising:

Using in Matching Model connection weight and biasing the user's characteristic information high dimension vector is weighted at summation Reason, user's characteristic information high dimension vector after being handled, the Matching Model is by being labelled with the training sample of true vector similarity This is trained；

Dimension-reduction treatment is carried out to user's characteristic information high dimension vector after the processing using the default dimensionality reduction function in Matching Model, Obtain user's characteristic information low-dimensional vector.

7. content recommendation method as claimed in claim 6, which is characterized in that using the connection weight in Matching Model and partially It sets and summation process is weighted to the user's characteristic information high dimension vector, user's characteristic information high dimension vector after being handled, Include:

When the connection weight in the Matching Model is nonnegative number, the user characteristics are believed according to connection weight and biasing Breath high dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled；

When the connection weight in the Matching Model is negative, according to the absolute value of connection weight and biasing to the user Characteristic information high dimension vector is weighted summation process, user's characteristic information high dimension vector after being handled.

8. content recommendation method as claimed in claim 6, which is characterized in that using the connection weight in Matching Model and partially Set and summation process be weighted to the user's characteristic information high dimension vector, after being handled user's characteristic information high dimension vector it Before, further includes:

Training sample is obtained, the training sample includes user's training sample and content training sample, the training sample mark True vector similarity between user's training sample and content training sample；

The initial matching model is restrained according to the true vector similarity and predicted vector similarity, is trained Matching Model afterwards.

9. content recommendation method as described in claim 1, which is characterized in that the user's characteristic information of target user is obtained, with And content DUAL PROBLEMS OF VECTOR MAPPING set of relationship, comprising:

Obtain the user's characteristic information of target user and the content characteristic information of content to be recommended, the content characteristic information Feature, content feature vector including content to be recommended；

The feature of the content to be recommended is mapped in default vector space, obtain the feature of content to be recommended to Amount；

The content feature vector and the feature vector of content to be recommended are subjected to Vector Fusion, obtain content to be recommended Content characteristic information vector；

According to the mapping relations content construction between the content to be recommended and the content characteristic information vector of content to be recommended DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

10. content recommendation method as claimed in claim 9, which is characterized in that the content characteristic information vector includes content The content feature vector and the feature vector of content to be recommended are carried out Vector Fusion, obtained by characteristic information high dimension vector To after the content characteristic information vector of content to be recommended, further includes:

According to the mapping relations content construction between the content to be recommended and the content characteristic information vector of content to be recommended DUAL PROBLEMS OF VECTOR MAPPING set of relationship, comprising:

According to the mapping relations building between the content to be recommended and the content characteristic information low-dimensional vector of content to be recommended Content DUAL PROBLEMS OF VECTOR MAPPING set of relationship.

11. such as the described in any item content recommendation methods of claim 1~10, which is characterized in that calculate the user characteristics letter Cease the vector similarity in vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship between content characteristic information vector, comprising:

The COS distance is normalized, the vector of user's characteristic information vector Yu content characteristic information vector is obtained Similarity.

12. a kind of content recommendation device, feature vector are, comprising:

Acquiring unit, for obtaining the user's characteristic information and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship of target user, the content DUAL PROBLEMS OF VECTOR MAPPING set of relationship includes the mapping relations between content to be recommended and the content characteristic information vector of content to be recommended；

It is corresponding to obtain the user's characteristic information for carrying out feature space mapping to the user's characteristic information for map unit User's characteristic information vector；

Computing unit, for calculating content characteristic information in the user's characteristic information vector and content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Vector similarity between vector；

Determination unit, for based on the vector similarity from content characteristic information vector determination object content characteristic information to Amount；

Selection unit, for choosing the corresponding mesh of object content characteristic information vector based on the content DUAL PROBLEMS OF VECTOR MAPPING set of relationship Mark content；

13. content recommendation device as claimed in claim 12, which is characterized in that the user's characteristic information includes user characteristics Content and user characteristics vector, the map unit include:

Map subelement, for by the user characteristics content map into default vector space, obtain user characteristics content to Amount；

Subelement is merged, for the user characteristics vector and user characteristics content vector to be carried out Vector Fusion, is obtained described The corresponding user's characteristic information vector of user's characteristic information.

14. content recommendation device as claimed in claim 13, which is characterized in that the mapping subelement is specifically used for:

15. a kind of server, which is characterized in that including processor and memory, the memory is stored with a plurality of instruction；It is described Processor loads instruction from the memory, to execute as in the described in any item content recommendation methods of claim 1~10 The step of.

16. a kind of storage medium, feature vector are, the storage medium is stored with a plurality of instruction, and described instruction is suitable for place Reason device is loaded, and requires the step in 1~10 described in any item content recommendation methods with perform claim.