CN113704617A

CN113704617A - Article recommendation method, system, electronic device and storage medium

Info

Publication number: CN113704617A
Application number: CN202111007695.2A
Authority: CN
Inventors: 刘亭
Original assignee: Kangjian Information Technology Shenzhen Co Ltd
Current assignee: Kangjian Information Technology Shenzhen Co Ltd
Priority date: 2021-08-30
Filing date: 2021-08-30
Publication date: 2021-11-26

Abstract

The invention provides an article recommendation method, a system, an electronic device and a storage medium, and relates to the technical field of machine learning, wherein the method comprises the following steps: acquiring article characteristics, and performing vectorization processing on the article characteristics to obtain an article vector; acquiring user characteristics, and performing vectorization processing on the user characteristics to obtain a user vector; carrying out vector retrieval processing on each user vector in a plurality of the article vectors through a preset vector retrieval tool, and retrieving a preset number of recommended article sets matched with each user vector; calculating cosine similarity between the articles and the user vectors in a user vector pair formed by each article in the recommended article set and the matched user vectors; and outputting an item recommendation list according to the sorting sequence of the cosine similarity. According to the method, the recommended article set is retrieved through the vector retrieval tool, and the article recommendation list is output according to cosine similarity sorting, so that the accuracy of the article recommendation list is improved.

Description

Article recommendation method, system, electronic device and storage medium

Technical Field

The invention relates to the technical field of machine learning, in particular to an article recommendation method, an article recommendation system, electronic equipment and a storage medium.

Background

With the rapid development of internet technology, people's lives are gradually becoming intelligent. When articles are browsed on a certain shopping platform, the shopping platform can collect historical behavior data of a user, corresponding features are extracted from the browsed articles according to the historical behavior data of the user, article recommendation is carried out according to the article features, and article recommendation is conveniently carried out according to user preferences so as to improve the experience of the user.

In the related art, a collaborative filtering algorithm is generally adopted for article recommendation, but the collaborative filtering algorithm has three defects, firstly, in the article recommendation process, for a new user, there is no historical behavior data, so that the system platform cannot capture the behavior preference of the user, and thus the articles which the user may be interested in cannot be matched. For new articles, the system platform cannot acquire the behavior of the user on the new articles, so that interested users cannot be matched with the new articles. Secondly, most users have sparse behavior data of the articles, so that the collaborative filtering algorithm consumes a large amount of resources during calculation, and the calculation result is inaccurate. Thirdly, the collaborative filtering algorithm can only capture the similarity between users with common user behaviors and the similarity between common article behaviors within a certain time, so that the accuracy of article recommendation is low.

Disclosure of Invention

The invention provides an article recommendation method, an article recommendation system, an electronic device and a storage medium, and mainly aims to improve article recommendation accuracy.

To achieve the above object, a first aspect of an embodiment of the present disclosure provides an item recommendation method, including:

acquiring article characteristics, and performing vectorization processing on the article characteristics to obtain an article vector;

acquiring user characteristics, and performing vectorization processing on the user characteristics to obtain a user vector;

carrying out vector retrieval processing on each user vector in a plurality of the article vectors through a preset vector retrieval tool, and retrieving a preset number of recommended article sets matched with each user vector;

calculating cosine similarity between the articles and the user vectors in a user vector pair formed by each article in the recommended article set and the matched user vectors;

and outputting an item recommendation list according to the sorting sequence of the cosine similarity.

In some embodiments, the outputting of the item recommendation list according to the cosine similarity ranking order comprises

And sequencing the cosine similarity from large to small to obtain a similarity sequence, and outputting an article recommendation list according to the similarity sequence.

In some embodiments, before obtaining the item feature, the item recommendation method further comprises:

acquiring user behavior data within a preset time range, wherein the user behavior data comprises characteristics of browsed articles;

substituting the browsed article characteristics into a preset interest matching model, and adjusting parameters of the preset interest matching model according to a preset error range to construct an article recommendation model, wherein the article recommendation model is used for vectorizing the article characteristics to obtain an article vector.

In some embodiments, the obtaining the article feature and performing vectorization processing on the article feature to obtain an article vector includes:

acquiring at least one of the following article characteristics: item identity tag information, item category, item brand, item title, item price;

converting the item features into a corresponding item feature sequence;

sequentially mapping the article characteristic sequences to a vector space;

and characterizing each article feature as a dimension of the article vector to obtain the article vector.

In some embodiments, the obtaining user features and vectorizing the user features to obtain a user vector includes:

at least one of the following user characteristics is obtained: user identity label information, a user click article sequence, user gender, user age, user mobile phone using parameters and user address information;

converting the user features into a corresponding user feature sequence;

sequentially mapping the user characteristic sequences to a vector space;

and characterizing each user feature as the dimension of the user vector to obtain the user vector.

In some embodiments, the vector retrieval tool is a Faiss vector retrieval tool, and the retrieving, by a preset vector retrieval tool, a preset number of recommended item sets matching each of the user vectors by performing vector retrieval processing for each of the user vectors in a plurality of the item vectors includes:

clustering the item vectors to determine a nearest cluster center point of each item vector;

acquiring a preset number of clustering center points closest to each user vector to obtain a clustering center set;

calculating the distance between each cluster center point in the cluster center set and the user vector;

and constructing a maximum heap of a preset number of elements for each user vector, putting the distance corresponding to the item vector into the maximum heap for heap sorting, and selecting the preset number of items from a sorting queue as the recommended item set of the user vector.

In some embodiments, the vector search tool is a Faiss vector search tool, and the vector search processing on the item vector and the user vector to search out a preset number of recommended item sets includes:

acquiring a preset number of clustering center points closest to the user vector to obtain a clustering center set;

and constructing a maximum heap of a preset number of elements, and putting the distance corresponding to the item vector into the maximum heap for heap sorting processing to obtain the recommended item set of the preset number.

In some embodiments, after outputting the item recommendation list according to the sorting order of the cosine similarity, the item recommendation method further comprises:

setting the recommended item set as a prediction sample;

acquiring user behavior data after the recommended article set is output, and extracting browsed articles according to the user behavior data to determine a real sample;

and calculating a loss function between the prediction sample and the real sample, and adjusting parameters of the item recommendation model according to the loss function.

To achieve the above object, a second aspect of the present disclosure proposes an item recommendation system, the system including:

the first vectorization module is used for acquiring article characteristics and vectorizing the article characteristics to obtain an article vector;

the second vector quantization module is used for acquiring user characteristics and carrying out vector processing on the user characteristics to obtain a user vector;

the retrieval module is used for carrying out vector retrieval processing on each user vector in a plurality of article vectors through a preset vector retrieval tool and retrieving a preset number of recommended article sets matched with each user vector;

the calculation module is used for calculating cosine similarity between the articles and the user vectors in a user vector pair formed by each article in the recommended article set and the matched user vectors;

and the output module outputs an article recommendation list according to the sorting sequence of the cosine similarity.

To achieve the above object, a third aspect of the present disclosure provides an electronic device, including:

at least one memory;

at least one processor;

at least one program;

the programs are stored in the memory, and the processor executes the at least one program to implement:

the method of the first aspect.

To achieve the above object, a fourth aspect of the present disclosure proposes a storage medium which is a computer-readable storage medium storing computer-executable instructions for causing a computer to perform:

the method of the first aspect.

According to the item recommendation method, the system, the electronic device and the storage medium, the item vectors and the user vectors are independently constructed, then vector retrieval processing is carried out on the plurality of item vectors aiming at the user vectors according to a preset vector retrieval tool to obtain a preset number of recommended item sets, cosine similarity between the items and the user vectors in the user vector pairs formed by each item in the recommended item sets and the matched user vectors is calculated after the recommended item sets are obtained, and then an item recommendation list is output according to the cosine similarity sorting sequence. The problem of cold start of the user and the articles is solved by independently constructing the article vector and the user vector, and the similarity of the articles on the context is also considered, so that the probability of recommending the similar articles can be reduced, the visual field of the user is improved, and the user experience is improved. The vector retrieval tool is used for retrieving the recommended article set and then cosine similarity calculation is carried out, cosine similarity calculation is not needed to be carried out on all article vectors and user vectors, and cosine similarity calculation time is saved, so that cosine similarity calculation efficiency is improved, and accuracy of the article recommendation list is improved.

Drawings

FIG. 1 is a flow chart of an item recommendation method provided by an embodiment of the present disclosure;

FIG. 2 is a flow chart of an item recommendation method provided by another embodiment of the present disclosure;

FIG. 3 is a flowchart of step S100 in FIG. 1;

FIG. 4 is a flowchart of step S200 in FIG. 1;

FIG. 5 is a flowchart of step S300 in FIG. 1;

FIG. 6 is a partial flow diagram of an item recommendation method provided by another embodiment of the present disclosure;

fig. 7 is a block diagram of an electronic device according to an embodiment of the invention.

The implementation, functional features and advantages of the objects of the present invention will be further explained with reference to the accompanying drawings.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.

It should be noted that although functional blocks are partitioned in a schematic diagram of an apparatus and a logical order is shown in a flowchart, in some cases, the steps shown or described may be performed in a different order than the partitioning of blocks in the apparatus or the order in the flowchart. The terms first, second and the like in the description and in the claims, and the drawings described above, are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to be limiting of the invention.

First, several terms related to the present application are analyzed:

vector retrieval: vector retrieval, an important way of computer intelligence retrieval. Each record (document representation or data entry) in the retrieval system is represented by a vector of weights of several index words, called a document vector. Clustering documents are generated by calculating the similarity between document vectors as a basis for retrieval. The questions (information requirements) of the user are also represented in the same way, called a question vector. The retrieval operation (i.e., matching operation of the question vector and the document vector) is performed in the clustered documents of the system. The similarity between a given question vector and a document (class) vector is calculated, and then documents with the similarity exceeding a certain threshold (or the number of documents to be detected according to the preset number) are sorted and output in descending order according to the similarity. The retrieval system adopting the mode realizes a local matching strategy and a sequencing output technology, and improves the flexibility and efficiency of retrieval.

Cosine similarity: cosine similarity, also called cosine similarity, is to evaluate the similarity of two vectors by calculating the cosine value of their included angle. Cosine similarity maps vectors into a vector space, such as the most common two-dimensional space, according to coordinate values. It is commonly used for file comparison in text mining. Furthermore, in the field of data mining, it is used to measure cohesion inside clusters.

Loss function (loss function): is a function that maps the value of a random event or its associated random variable to a non-negative real number to represent the "risk" or "loss" of the random event. In application, the loss function is usually associated with the optimization problem as a learning criterion, i.e. the model is solved and evaluated by minimizing the loss function. For example, in statistics and machine learning, parameter estimation (parameter estimation) of models, in macro-economics, risk management (risk management) and decision-making, and in control theory, optimal control theory (optimal control theory).

Embedding (embedding): embedding is a vector representation, which means that a low-dimensional vector represents an object, which can be a word, a commodity, a movie, etc.; the embedding vector has the property that objects corresponding to vectors with similar distances have similar meanings, for example, the distance between the embedding (revenge league) and the embedding (ironmen) is very close, but the distance between the embedding (revenge league) and the embedding (dinners) is far away. The embedding essence is mapping from a semantic space to a vector space, and simultaneously, the relation of an original sample in the semantic space is kept as much as possible in the vector space, for example, the positions of two words with similar semantics in the vector space are also relatively close. The embedding can encode an object by using a low-dimensional vector and also can reserve the meaning of the object, is usually applied to machine learning, and in the process of constructing a machine learning model, the object is encoded into a low-dimensional dense vector and then transmitted to the DNN, so that the efficiency is improved.

The important components of the item recommendation method comprise: recall algorithms and recall strategies, the quality of the recall largely determining the quality of the recommendation. The recall algorithm adopted by the current recommendation method is mainly divided into a recall based on collaborative filtering and a recall based on deep learning. Among them, there are three problems with collaborative filtering based recall, first, the cold start problem, second, the data sparsity problem, and third, the scalability problem. For the new article, the algorithm cannot capture the behavior preference of the user due to no historical behavior, so that the article possibly interested by the user cannot be matched. For the problem of data sparsity, a scene generally faced by the recommendation method is a set formed by large-scale users and articles, the set is continuously expanded, behaviors of most users to the articles are sparse, so that a collaborative filtering algorithm consumes a large amount of resources during calculation, and the accuracy of a result of recommending the articles is low. Aiming at the problem of expandability, the collaborative filtering algorithm can only capture the similarity between the common user behaviors and the users and the similarity between the articles with the common article behaviors in a certain time, and cannot capture the semantic or emotional similarity and the correlation between the articles, for example, in the sports shoes and the sports coats with the behaviors not intersecting, the collaborative filtering algorithm cannot calculate the similarity between the two, but in the context, the sports shoes and the sports coats belong to the sports system and have the similarity on the context, so that the articles are recommended through the collaborative filtering algorithm, the similar articles are easily and more recommended, the visual field of the user is narrower and narrower, and the user experience is influenced.

The embodiment of the application can acquire and process related data based on an artificial intelligence technology. Among them, Artificial Intelligence (AI) is a theory, method, technique and application system that simulates, extends and expands human Intelligence using a digital computer or a machine controlled by a digital computer, senses the environment, acquires knowledge and uses the knowledge to obtain the best result.

The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a robot technology, a biological recognition technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

Based on the above, the application discloses an article recommendation method, which includes the steps of respectively processing according to user characteristics and article characteristics to obtain corresponding user vectors and article vectors, then obtaining a recommended article set related to a user by adopting a vector retrieval tool, then calculating cosine similarity between articles and the user vectors in a user vector pair formed by each article and a matched user vector, and outputting an article recommendation list according to the cosine similarity so as to improve article recommendation precision.

The article recommendation method, the system, the electronic device and the storage medium provided by the embodiments of the present disclosure are specifically described by the following embodiments, and first, the article recommendation method in the embodiments of the present disclosure is described.

Fig. 1 is an optional flowchart of an item recommendation method provided in an embodiment of the present disclosure, and the method in fig. 1 may include, but is not limited to, steps S100 to S500:

s100, acquiring article characteristics, and performing vectorization processing on the article characteristics to obtain an article vector;

s200, acquiring user characteristics, and performing vectorization processing on the user characteristics to obtain a user vector;

s300, performing vector retrieval processing on each user vector in a plurality of article vectors through a preset vector retrieval tool, and retrieving a preset number of recommended article sets matched with each user vector;

s400, calculating cosine similarity between the articles and the user vectors in the user vector pair formed by each article in the recommended article set and the matched user vectors;

and S500, outputting an item recommendation list according to the sorting sequence of the cosine similarity.

The method comprises the steps of obtaining article features and user features, conducting vectorization processing on the article features to obtain article vectors, conducting vectorization processing on the user features to obtain user vectors, and conducting vector retrieval on the user vectors in the article vectors through a preset vector retrieval tool to obtain a preset number of recommended article sets. Before cosine similarity calculation between each article and each user vector pair is carried out, a vector retrieval tool is adopted to retrieve recommended article sets with preset quantity, similarity between the user vectors and each article vector does not need to be directly calculated, the recommended article sets with the preset quantity related to the user vectors are retrieved first, and then the cosine similarity between the articles in the user vector pairs and the user vectors in the user vector pairs formed by each article in the recommended article sets and the matched user vectors is calculated, so that the cosine similarity calculation of the article and the user vector pairs is simplified, and the article recommendation efficiency is improved.

The article feature is vectorized to obtain an article vector, the user feature is vectorized to obtain a user vector, the article vector is article embedding, and the user vector is user embedding. The embedding is a way of converting discrete variables into continuous variable representations, that is, is used for visualizing the relationship between different discrete variables, and the embedding can reduce the spatial dimension of the discrete variables and can also represent the variables meaningfully. Therefore, the article features are subjected to vectorization processing, that is, discrete article features are converted into continuous variables, then the article features are represented through article embedding, and the user features are subjected to vectorization processing, that is, discrete user features are converted into continuous variables, so that the user features are represented through user embedding. The article vector and the user vector are calculated independently, a preset number of recommended article sets are retrieved through a vector retrieval tool, namely, the correlation between the user and the article is calculated according to the user vector and the article vector, and the calculation amount of cosine similarity between the article and the user vector in the user vector pair is saved by adopting a vector distance measurement mode, so that the article recommendation efficiency is improved.

In some embodiments, referring to fig. 2, before performing step S100, the method for recommending an item further includes:

s600, obtaining user behavior data in a preset time range, wherein the user behavior data comprises characteristics of browsed articles;

s700, substituting the characteristics of the browsed articles into a preset interest matching model, and adjusting the parameters of the preset interest matching model according to a preset error range to construct an article recommendation model, wherein the article recommendation model is used for vectorizing the characteristics of the articles to obtain an article vector.

The method comprises the steps of obtaining user behavior data within a preset time range, wherein the obtained user behavior data mainly comprises the steps that a user clicks an article on a system plane for browsing, inputting a preset interest matching model through characteristics of the article browsed in the user behavior data, adjusting parameters of a preset interest module according to a preset error range to construct an article recommendation model, and the article recommendation model also calculates the similarity between two articles or between two users to obtain a recommended article. The method comprises the steps of setting an article training feature and an article training vector according to browsed article features, enabling the article training vector to represent the features of an article through a digital vector, inputting the article training feature into a preset interest matching model, adjusting parameters of the preset interest matching model according to the article training vector and a preset error range, and controlling a difference value between the article vector and the article training vector output by the preset interest matching model to be within the preset error range to obtain an article recommendation model. The item recommendation model is established according to the items browsed by the user, the item recommendation models corresponding to different users are different, so that the item vector which accords with the user behavior data is calculated, the recommended items calculated according to the item recommendation model are wider, and the experience of the user is improved. The preset time range is customized according to users, so that corresponding preset time ranges can be set according to different users, user behavior data in corresponding time areas are collected, and a corresponding article recommendation model is built according to the user behavior data in the preset time ranges. Therefore, for a user with user behavior data, the continuously updated articles are directly substituted into the article recommendation model for vectorization calculation to calculate the similarity between the articles, so that the article recommendation range is expanded, the probability of similar article recommendation is reduced, the visual field of the user is improved, and the user experience is improved.

For example, the preset time range is set by self-defining mainly according to the time of the user using the system, if user behavior data exist for a long time, the preset time range is set to be 30 days, and for a new user, the time range of the user behavior data existing is smaller than 30 days, the user behavior data of the user using interval are obtained, and the preset time range is continuously modified according to the using duration of the user, so that the collecting time range of the user behavior data is gradually expanded. After the preset time range is determined, user behavior data in the preset time range are collected according to the preset time interval period, and the article recommendation model is updated according to the preset time interval period, so that the accuracy of the article recommendation model is improved. For example, if the preset time interval is 1 day and the preset time range is 30 days, the user behavior data of the previous 30 days is collected at the 31 st day, and the user behavior data of the previous 30 days is collected again at the 32 nd day, so as to update the item recommendation model in real time according to the user behavior data.

Specifically, the preset interest matching model is a DSSM model, the DSSM model is also called a deep semantic matching model, and a user of the DSSM model predicts semantic similarity of two sentences, and can also obtain a low-dimensional semantic vector expression of a certain sentence, so as to calculate similarity between features of two articles through the DSSM model, that is, determine similarity between two articles through an article vector. The DSSM model can be divided into three layers, the three layers are an input layer, a presentation layer and a matching layer, the input layer maps the features into a vector space and inputs the vector space into DNN, the presentation layer is equivalent to discarding the position information of the word vector, namely all the features are placed in a bag, and the sequence is not required. The matching layer can also calculate the cosine similarity of the two semantic vectors. Therefore, the similarity between the characteristics of different articles is calculated through the DSSM model. Therefore, the article features are substituted into the article recommendation model to obtain an article vector, the similarity between the two articles is represented by the article vector, and the article vector has the similarity on the context, so that the article recommendation effect is reduced, the article recommendation range is expanded, and more articles according with the preference of the user can be obtained.

Referring to fig. 3, in some embodiments, step S100 includes, but is not limited to, steps S110 to S140:

s110, at least one of the following article characteristics is obtained: item identity tag information, item category, item brand, item title, item price;

s120, converting the article characteristics into corresponding article characteristic sequences;

s130, sequentially mapping the article characteristic sequences to a vector space;

s140, characterizing each article feature as the dimension of the article vector to obtain the article vector.

As for the new article, the behavior of the user on the new article does not exist, so that similar articles and interested users cannot be matched with the new article, all article characteristics are subjected to vectorization processing, and each article vector can be obtained only according to the article characteristics. The similarity between the articles is represented through the article vectors, the article vectors corresponding to each article are independently constructed, and the article identity label information, the article categories, the article brands, the article titles and the article prices are used as the dimensions of the article vectors, so that the characteristic parameters of each article can be clear according to the obtained article vectors. Since the article features are not in a digital form, the article features are converted into article feature sequences, then the article feature sequences are sequentially mapped to a vector space, and the article features are taken as the dimensions of an article vector to obtain the article vector, so that the article features are represented in a digital form.

For example, the article identification tag information is also the article ID, and the article ID corresponding to each article is unique; the item category then represents the type of item, and the item category includes: washing products, clothes, mother and baby products, furniture, kitchen products, footwear and the like to distinguish each item according to the category of the item; the article brands are a plurality of article brands according to the same category, for example, the washing and protecting articles comprise a plurality of brands such as 'union lihua' and 'Baojie'; the title of the article is a specific name of the article, such as 'Haifeisi dandruff-removing shampoo'; the price of the item is the selling price of the item. Therefore, by using the item ID, the item category, the item brand, the item title, and the item price as the dimensions of the item vector and representing the features corresponding to each item through the item vector, the recommended item set calculated according to the item vector better conforms to the user preference, so as to improve the user experience.

Referring to fig. 4, in some embodiments, step S200 includes, but is not limited to, steps S210 to S240:

s210, at least one of the following user characteristics is obtained: user identity label information, a user click article sequence, user gender, user age, user mobile phone using parameters and user address information;

s220, converting the user characteristics into a corresponding user characteristic sequence;

s230, sequentially mapping the user characteristic sequences to a vector space;

and S240, representing each user feature as the dimension of the user vector to obtain the user vector.

Because the vector retrieval tool retrieves a preset number of recommended item sets according to the item vectors and the user vectors, in order to make the recommended item sets more in line with the user preference, each user feature is used as the dimension of the user vector by acquiring the user identity tag information, the user click item sequence, the user gender, the user using mobile phone parameters and the user address information, and the user vector is used for representing the user feature, vector retrieval processing can be performed according to the user vector and the item vectors, so that the obtained recommended item sets more in line with the user preference, and the user experience is improved. The user features are characters, the user feature sequences need to be converted into corresponding user feature sequences, then the user feature sequences are sequentially mapped into the vectors, the user features are used as the dimensionalities of the user vectors to obtain the user vectors, the user features are represented in a digital form, and therefore the user favorite article recommendation list can be found conveniently according to the user vectors and the article vectors.

For example, the user identity tag information is a user ID; collecting the items clicked by the user within a preset time range if the user clicks the item sequence, and carrying out time sequencing according to the items clicked by the user to obtain an item click sequence; the gender of the user is determined according to the registered gender of the user during registration; the age of the user is determined according to the month of birth year registered by the user; the parameters of the mobile phone used by the user are directly read, and the parameters of the mobile phone used by the user comprise: user mobile phone brand, user mobile phone model, user mobile phone specification. The user address information includes: the system comprises a user real-time address and a user registration address, wherein the user real-time address is determined according to an IP address of each login of a user, and the user registration address is determined according to address information filled in by the user during registration. Therefore, by using the user identity tag information, the user click item sequence, the user gender, the user age, the user use mobile phone parameters and the user address information as the dimensionality of the user vector, the characteristics of each user can be clarified according to the user vector, and the recommended item set retrieved according to the user vector and the item vector also better conforms to the user preference so as to improve the user experience.

In some embodiments, step S300 may include, but is not limited to including:

by at least one of the following vector retrieval tools: and the Faiss vector retrieval tool and the SPTAG vector retrieval tool are used for carrying out vector retrieval processing on the article vectors and the user vectors and retrieving a preset number of recommended article sets.

The vector retrieval tool mainly carries out vector retrieval on all the article vectors, retrieves the article vectors with the preset number nearest to the user vector, and obtains a recommended article set. The recommended item set is retrieved through the vector retrieval tool, all item vectors and user vectors do not need to be collected when cosine similarity calculation is carried out, and only the cosine similarity of the item vectors and the user vectors in the recommended item set needs to be calculated, so that the power consumption of the cosine similarity calculation is saved, and the speed of the cosine similarity calculation is increased.

The item vectors and the user vectors are used for representing item features and user features by numbers, so that the matching degree between a user and each item is calculated, namely vector retrieval processing is carried out on the user vectors and the item vectors through a vector retrieval tool, namely the distance between the user vectors and the item vectors is calculated, the similarity between the item vectors and the user vectors is judged according to the distance, and then the item vectors with the preset number closest to the user vectors are screened out. Wherein the vector retrieval tool comprises: the device comprises a Faiss vector retrieval tool and an SPTAG vector retrieval tool, wherein the Faiss vector retrieval tool is an open-source algorithm library which is written by a Facebook AI based on a C + + language and aims at multimedia file similarity search. The Faiss vector retrieval tool supports the optimization setting of a developer on retrieval speed, memory use, retrieval precision and the like, but is only an algorithm library and has higher use requirements on the developer. The SPTAG vector search tool uses a graph-based nearest neighbor search algorithm from a library of vector search algorithms published by Microsoft. The SPTAG vector retrieval tool has the advantages of high search speed, intelligent search of billions of vectors within milliseconds, good performance in query accuracy and memory occupation, obvious defects, long graph building time, and the need of re-building a graph each time a new vector is added into a database. In the embodiment, a Faiss vector retrieval tool is adopted to calculate the item vector and the user vector so as to retrieve a preset number of recommended item sets, and the Faiss vector retrieval tool supports vector-based recall, graph search and the like, so that the method is suitable for being applied to a recall algorithm for item recommendation.

Referring to fig. 5, in some embodiments, the vector search tool is a Faiss vector search tool, and when the vector search process is performed by the Faiss vector search tool, step S300 further includes, but is not limited to, steps S310 to S340:

s310, clustering the article vectors to determine the nearest clustering center point of each article vector;

s320, obtaining a preset number of clustering center points closest to the user vector to obtain a clustering center set;

s330, calculating the distance between each cluster center point in the cluster center set and the user vector;

s340, constructing a maximum heap of a preset number of elements for each user vector, putting the distance corresponding to the item vector into the maximum heap for heap sorting, and selecting the preset number of items from the sorting queue as a recommended item set of the user vector.

The method comprises the steps of carrying out vector retrieval processing by adopting a Faiss vector retrieval tool, firstly carrying out clustering processing on each article vector to obtain a clustering center closest to each article vector, determining a clustering center point of the article vector due to the fact that the article vector is multi-dimensional, then obtaining a preset number of clustering center points closest to a user vector to obtain a clustering center set, then calculating the distance between the clustering center point in the clustering center set and the user vector to represent the matching degree of each article vector and the user vector through the distance, then placing the distance corresponding to each article vector into a maximum heap to carry out heap sorting processing, and selecting a preset number of articles from a sorting queue to be used as a recommended article set. Therefore, by finding the clustering center point corresponding to the item vector, calculating the distance between the user vector and the clustering center point, and obtaining the preset number of item vectors according to the distance sequence to obtain the recommended item set, the recommended item set is more in line with the preference of the user, so that the experience of the user is improved.

Specifically, after the recommended item set is obtained, an item and user vector pair needs to be determined according to the recommended item set, and then cosine similarity between each item and the user vector pair is calculated, so that it is clear that each item better meets user preferences according to the cosine similarity, and an item recommendation list is output, so that user experience is improved.

For example, the cosine similarity calculation mainly uses a cosine value of an included angle between two vectors in a vector space as a measure for measuring the difference between two individuals, the closer the cosine value is to 1, which indicates that an article vector is closer to a user vector, and the cosine similarity calculation formula is as follows:

if the user vector is: (1, 2, 1, 0, 1), the item vector is: (1, 1, 3, 0, 1), calculating the cosine similarity of the user vector and the item vector as follows:

therefore, the cosine similarity of the user vector and the item vector is calculated to characterize the matching degree of the items and the preference of the user according to the cosine similarity, similarity sorting can be performed according to the cosine similarity corresponding to each item, and the item recommendation list output according to the similarity sorting is more in line with the preference of the user, so that the experience of the user is improved.

In some embodiments, step S500 may include, but is not limited to including:

The cosine similarity is sorted from large to small to obtain a similarity sequence, the similarity sequence is also the sorting sequence of articles favored by the user, then the article recommendation list is output according to the similarity sequence, sorting output can be carried out according to the preference of the user to the articles, and the user can quickly acquire the articles to be looked up according to the article recommendation list to improve the experience of the user.

Referring to fig. 6, in some embodiments, after step S500, the item recommendation method further includes, but is not limited to, the steps of:

s800, setting the recommended item set as a prediction sample;

s900, obtaining user behavior data after the recommended article set is output, and extracting browsed articles according to the user behavior data to determine a real sample;

s1000, calculating a loss function between the prediction sample and the real sample, and adjusting parameters of the item recommendation model according to the loss function.

In order to improve the accuracy of the item recommendation model, items need to be extracted and browsed according to user behavior data to obtain real samples, then the item recommendation set is taken as a prediction sample to calculate a loss function between the real samples and the prediction sample, and then the item recommendation model is continuously optimized according to the loss function to improve the accuracy of an item vector calculated according to the item recommendation model, so that the recommended item set obtained by vector retrieval processing according to the item vector and the user vector better conforms to the preference of the user.

The loss function is calculated for the prediction sample and the real sample, and the two classification processing is also carried out for the prediction sample and the real sample to obtain the loss function, wherein the loss function is a cross entropy loss function. If the real sample is y, the prediction sample is x, then a sample point (x, y) is constructed according to the prediction sample and the real sample, y represents the real sample, and the value of y is 0 or 1, that is, whether the real sample and the prediction sample are the same is judged through the value of y, if so, y is 1, otherwise, y is 0. Suppose that the true sample of a certain sample point is y_tThe real sample is taken_tProbability of 1 being y_pThen the loss function for that sample point is:

-log(y_t|y_p)＝-(ylog(y_p)+(1-y_t)log(1-y_p))

the loss function is also used for calculating the matching degree of the recommended article obtained by prediction and the article actually liked by the user, and the loss function is used as a reference for adjusting parameters of the article recommendation model, so that the loss function is gradually close to 1, namely the recommended article set is shown to be in line with the preference of the user, and the accuracy of the article recommendation set is improved.

An embodiment of the present disclosure further provides an article recommendation system, which may implement the article recommendation method, where the system includes:

the first vectorization module is used for acquiring the article characteristics and vectorizing the article characteristics to obtain an article vector;

the second vector quantization module is used for acquiring the user characteristics and carrying out vector processing on the user characteristics to obtain a user vector;

the retrieval module is used for carrying out vector retrieval processing on each user vector in the multiple article vectors through a preset vector retrieval tool and retrieving a preset number of recommended article sets matched with each user vector;

and the output module outputs the item recommendation list according to the sorting sequence of the cosine similarity.

The method comprises the steps of acquiring article features, vectorizing the article features to obtain article vectors, vectorizing the acquired user vectors to obtain user vectors, representing the article features through the article vectors independently, representing the user features through the user vectors, then carrying out vector retrieval processing on a plurality of article vectors and the user vectors according to a vector retrieval tool to retrieve recommended article sets in a preset number, retrieving the recommended article sets through the vector retrieval tool, and only calculating cosine similarity between articles in the recommended article sets and the user vectors in user vector pairs formed by matched user vectors and each article in the recommended article sets one by one without calculating the cosine similarity between the articles and the user vectors, so that the cosine similarity is calculated more quickly. After the cosine similarity corresponding to each article vector is obtained, the cosine similarities are sequenced to obtain similarity sequencing, then an article recommendation list is determined according to the similarity sequencing, and the article recommendation list is output, so that the article recommendation list obtained through calculation better accords with the preference of a user, the problem of cold start of the user and articles is solved by constructing the article vectors and the user vectors, the more accurate article recommendation list is obtained, and the user experience is improved.

An embodiment of the present disclosure further provides an electronic device, including:

at least one memory;

at least one processor;

at least one program;

the programs are stored in the memory and the processor executes the at least one program to implement the present disclosure to implement the item recommendation method described above. The electronic device can be any intelligent terminal including a mobile phone, a tablet computer, a Personal Digital Assistant (PDA for short), a vehicle-mounted computer and the like.

Referring to fig. 7, fig. 7 illustrates a hardware structure of an electronic device according to another embodiment, where the electronic device includes:

the processor 110 may be implemented by a general-purpose CPU (central processing unit), a microprocessor, an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits, and is configured to execute a relevant program to implement the technical solution provided by the embodiment of the present disclosure;

the memory 120 may be implemented in a ROM (read only memory), a static memory device, a dynamic memory device, or a RAM (random access memory). The memory 120 may store an operating system and other application programs, and when the technical solution provided by the embodiments of the present disclosure is implemented by software or firmware, the relevant program codes are stored in the memory 120 and called by the processor 110 to execute the item recommendation method according to the embodiments of the present disclosure;

an input/output interface 130 for implementing information input and output;

the communication interface 140 is configured to implement communication interaction between the device and other devices, and may implement communication in a wired manner (e.g., USB, network cable, etc.) or in a wireless manner (e.g., mobile network, WIFI, bluetooth, etc.); and

a bus 150 that transfers information between various components of the device (e.g., the processor 110, the memory 120, the input/output interface 130, and the communication interface 140);

wherein the processor 110, memory 120, input/output interface 130, and communication interface 140 are communicatively coupled to each other within the device via bus 150.

The embodiment of the disclosure also provides a storage medium which is a computer-readable storage medium, and the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used for causing a computer to execute the item recommendation method.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, apparatus, article, or method. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, apparatus, article, or method that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments. Through the above description of the embodiments, those skilled in the art will clearly understand that the method of the above embodiments can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware, but in many cases, the former is a better implementation manner. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium (e.g., ROM/RAM, magnetic disk, optical disk) as described above and includes instructions for enabling a terminal device (e.g., a mobile phone, a computer, a server, or a network device) to execute the method according to the embodiments of the present invention.

The above description is only a preferred embodiment of the present invention, and not intended to limit the scope of the present invention, and all modifications of equivalent structures and equivalent processes, which are made by using the contents of the present specification and the accompanying drawings, or directly or indirectly applied to other related technical fields, are included in the scope of the present invention.

Claims

1. An item recommendation method, characterized in that the method comprises:

2. The item recommendation method according to claim 1, wherein said outputting an item recommendation list according to said cosine similarity sorting order comprises

3. The item recommendation method according to claim 1, wherein before obtaining the item feature, the item recommendation method further comprises:

4. The item recommendation method according to claim 3, wherein the obtaining item features and vectorizing the item features to obtain an item vector comprises:

converting the item features into a corresponding item feature sequence;

sequentially mapping the article characteristic sequences to a vector space;

5. The item recommendation method according to claim 3, wherein the obtaining user characteristics and vectorizing the user characteristics to obtain a user vector comprises:

converting the user features into a corresponding user feature sequence;

sequentially mapping the user characteristic sequences to a vector space;

6. The item recommendation method according to any one of claims 1 to 5, wherein the vector search tool is a Faiss vector search tool, and the searching a set of recommended items matching each of the user vectors by performing a vector search process for each of the user vectors in the plurality of item vectors using a predetermined vector search tool includes:

7. The item recommendation method according to claim 3, wherein after outputting an item recommendation list according to the sorting order of the cosine similarity, the item recommendation method further comprises:

setting the recommended item set as a prediction sample;

8. An item recommendation system, the system comprising:

9. An electronic device, comprising:

at least one memory;

at least one processor;

at least one program;

the method of any one of claims 1 to 7.

10. A storage medium that is a computer-readable storage medium having stored thereon computer-executable instructions for causing a computer to perform:

the method of any one of claims 1 to 7.