Personalized product recommendation method, system, electronic device and readable storage medium

Abstract

The present invention relates to the field of data analysis technology, and discloses a method, system, electronic device, and readable storage medium for personalized product recommendation. The method obtains a user embedding vector representing user interests and a product embedding vector representing candidate products, and obtains a user feedback vector by mapping the user embedding vector and the product embedding vector. The user embedding vector and the user feedback vector are refined to obtain a vector refined representation, and then the user interest representation in the vector refined representation is calculated to obtain a related interest representation. A recommendation feature vector is generated based on the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation. The user click probability of the candidate product is predicted based on the recommendation feature vector, thereby determining the recommendation target, avoiding the noise impact on the candidate product caused by sparse historical data or cross-domain data heterogeneity, and improving the fit between the recommended product and the user's interest.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a method and system for personalized product recommendation, an electronic device, and a readable storage medium. Background Art

[0002] With the rapid development of internet technology and the explosive growth of information, users have greater access to a wide range of information. However, faced with a vastly overwhelmed information processing capacity, users are forced to spend considerable time filtering out content that is uninteresting or irrelevant to them. Providing personalized recommendations based on user history is one approach to addressing this issue. For example, recommending e-commerce products based on a user's shopping history can be helpful.

[0003] However, for vehicle purchasing platforms, most users have relatively sparse purchase records and lack historical interaction data, making it difficult to grasp user interests. As a result, the recommended products do not match user interests, reducing the user experience of personalized recommendations. Summary of the Invention

[0004] In order to provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. The summary is not an extensive review, nor is it intended to identify key / critical elements or delineate the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.

[0005] In view of the above-mentioned shortcomings of the prior art, the present invention discloses a personalized product recommendation method, system, electronic device and readable storage medium to improve the fit between recommended products and user interests.

[0006] The present invention provides a method for personalized product recommendation, comprising: obtaining a user embedding vector and a product embedding vector, wherein the user embedding vector is used to represent the interest information of a target user, and the product embedding vector is used to represent the product information of a candidate product; mapping the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and calculating the user embedding vector and the user feedback vector based on a self-attention mechanism to obtain a vector refined representation; extracting a user interest representation from each of the vector refined representations, and calculating based on the extracted user interest representations to obtain a related interest representation corresponding to each of the vector refined representations; calculating a recommendation feature vector based on a preset probability prediction model to obtain a user click probability of the candidate product, and determining a recommendation target from the candidate product based on the user click probability, wherein the recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation.

[0007] Optionally, obtaining a user embedding vector and a product embedding vector includes: obtaining interest information of a target user and product information of a candidate product; embedding the interest information and the product information into low-dimensional dense vectors, respectively, to obtain a user embedding vector corresponding to the interest information and a product embedding vector corresponding to the product information.

[0008] Optionally, after obtaining the target user's interest information and the product information of the candidate products, and before embedding the interest information and the product information into low-dimensional dense vectors respectively, the method further includes: performing data processing on the interest information and the product information, wherein the data processing includes one or more of data cleaning, data completion, and data standardization.

[0009] Optionally, obtaining a user feedback vector based on mutual mapping of the user embedding vector and the product embedding vector includes: establishing an orthogonal mapping model, wherein the orthogonal mapping model includes a first mapping model and / or a second mapping model; iteratively updating the orthogonal mapping model based on a dual learning mechanism, and determining an iterative loss value; determining that the iterative update of the orthogonal mapping model is completed when the iterative loss value converges; orthogonally mapping the user embedding vector to the product embedding vector through the first mapping model to obtain a first feedback vector, and orthogonally mapping the product embedding vector to the user embedding vector through the second mapping model to obtain a second feedback vector; and determining the first feedback vector and / or the second feedback vector as the user feedback vector.

[0010] Optionally, the user embedding vector and the user feedback vector are calculated respectively based on the self-attention mechanism to obtain a vector refined representation: a first query matrix, a first key matrix and a first value matrix of the vector to be refined are calculated, wherein the vector to be refined includes the user embedding vector or the user feedback vector; calculation is performed according to the first query matrix, the first key matrix and the first value matrix to obtain a first attention head output corresponding to the vector to be refined; the first attention head outputs are spliced ​​to obtain a multi-head attention output; the multi-head attention output is sequentially subjected to feedforward network processing, residual link, degradation layer and normalization processing to obtain a vector refined representation corresponding to the vector to be refined.

[0011] Optionally, extracting a user interest representation from each of the vector refined representations includes: calculating a second query matrix, a second key matrix, and a second value matrix of the vector refined representation; performing calculations based on the second query matrix, the second key matrix, and the second value matrix to obtain a second attention head output; and using the second attention head output as the user interest representation of the vector refined representation.

[0012] Optionally, calculations are performed based on the extracted user interest representations to obtain relevant interest representations corresponding to each of the vector refined representations, including: calculations are performed based on the product embedding vectors and the user interest representations to obtain first weights corresponding to each of the user interest representations; and weighted calculations are performed on the user interest representations based on the first weights to obtain relevant interest representations corresponding to the vector refined representations.

[0013] Optionally, a recommendation feature vector is generated by at least one of the following methods: concatenating the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation to obtain a recommendation feature vector; respectively calculating the second weights of the vector refined representation and the related interest representation, and fusing the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation according to the second weights to obtain a recommendation feature vector.

[0014] The present invention provides a personalized product recommendation system, comprising: an acquisition module for acquiring a user embedding vector and a product embedding vector, wherein the user embedding vector is used to represent the interest information of a target user, and the product embedding vector is used to represent the product information of a candidate product; a mapping module for mapping the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and calculating the user embedding vector and the user feedback vector based on a self-attention mechanism to obtain a vector refined representation; an extraction module for extracting a user interest representation from each of the vector refined representations, and calculating based on the extracted user interest representations to obtain a related interest representation corresponding to each of the vector refined representations; a determination module for calculating a recommendation feature vector based on a preset probability prediction model to obtain a user click probability of the candidate product, and determining a recommendation target from the candidate product based on the user click probability, wherein the recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation.

[0015] The present invention provides an electronic device, comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the above method.

[0016] The present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method described above is implemented.

[0017] Beneficial effects of the present invention:

[0018] This approach obtains a user embedding vector representing user interests and an item embedding vector representing candidate items. Mapping the user embedding vector and item embedding vector yields a user feedback vector, which is then refined to obtain a vector-refined representation. The user interest representation in the vector-refined representation is then calculated to obtain a related interest representation. A recommendation feature vector is generated based on the user embedding vector, item embedding vector, vector-refined representation, and related interest representation. The user click probability of the candidate item is then predicted based on the recommendation feature vector, thereby determining the recommendation target. In this way, user interest information is transferred to candidate item information through inter-domain mapping, and the vector is refined using a self-attention operation to obtain a vector-refined representation. The click probability of the candidate item is then predicted based on the user interest in the vector-refined representation. Compared to personalized recommendations based on historical shopping data for current purchases, this approach avoids the noise impact on candidate items caused by sparse historical data or heterogeneous cross-domain data, thereby improving the fit between recommended items and user interests. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flowchart of a personalized product recommendation method according to an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of a system framework for implementing a personalized product recommendation method in an embodiment of the present invention;

[0021] Figure 3 Schematic diagram of the mapping between the A domain embedding vector and the B domain embedding vector in an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of the structure of a personalized product recommendation system according to an embodiment of the present invention;

[0023] Figure 5 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and sub-samples in the embodiments can be combined with each other unless there is a conflict.

[0025] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0026] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0027] In the description and claims of the embodiments of the present disclosure, as well as in the accompanying drawings, the terms "first," "second," and the like are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate to describe the embodiments of the present disclosure herein. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions.

[0028] Unless otherwise stated, the term "plurality" means two or more.

[0029] In the embodiment of the present disclosure, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B means: A or B.

[0030] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0031] Combine Figure 1 As shown, the embodiment of the present disclosure provides a method for personalized product recommendation, including:

[0032] Step S101, obtaining user embedding vectors and product embedding vectors;

[0033] Among them, the user embedding vector is used to represent the target user's interest information, such as user age, user education, user income, user's historical purchase records on different platforms, etc.

[0034] Among them, the product embedding vector is used to represent the product information of the candidate product, such as vehicle price, vehicle appearance, vehicle configuration, etc.

[0035] Step S102: Mapping the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and calculating the user embedding vector and the user feedback vector based on the self-attention mechanism to obtain a vector refined representation;

[0036] Step S103: extracting user interest representations from each vector refined representation, and performing calculations based on the extracted user interest representations to obtain relevant interest representations corresponding to each vector refined representation;

[0037] In step S104, the recommendation feature vector is calculated according to a preset probability prediction model to obtain the user click probability of the candidate product, and the recommendation target is determined from the candidate products based on the user click probability. The recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation.

[0038] The disclosed embodiments provide a method for personalized product recommendation. This method obtains a user embedding vector representing user interests and a product embedding vector representing candidate products, maps the user embedding vector and the product embedding vector to obtain a user feedback vector, refines the user embedding vector and the user feedback vector to obtain a vector refined representation, and then calculates the user interest representation in the vector refined representation to obtain a related interest representation. A recommendation feature vector is generated based on the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation. The user click probability of the candidate product is predicted based on the recommendation feature vector, thereby determining the recommendation target. In this way, user interest information is transferred to candidate product information through inter-domain mapping, and the vector is refined by performing a self-attention operation to obtain a vector refined representation. The click probability of the candidate product is predicted based on the user interest in the vector refined representation. Compared to personalized recommendations for current purchases based on historical shopping data, this method avoids the noise impact on candidate products caused by sparse historical data or heterogeneous cross-domain data, thereby improving the fit between recommended products and user interests.

[0039] like Figure 2As shown, the embodiment of the present disclosure provides a system framework for implementing a personalized product recommendation method, including a dual embedding layer, an interest learning layer, a dual interest matching layer, and a prediction layer; wherein the dual embedding layer is used to obtain a user embedding vector and a product embedding vector, and map the user embedding vector and the product embedding vector to each other to obtain a user feedback vector; the interest learning layer is used to map the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and calculate the user embedding vector and the user feedback vector respectively based on the self-attention mechanism to obtain a vector refined representation; the dual interest matching layer is used to extract user interest representations from each vector refined representation, and calculate based on the extracted user interest representations to obtain relevant interest representations corresponding to each vector refined representation; the prediction layer is used to calculate the recommendation feature vector according to a preset probability prediction model to obtain the user click probability of the candidate product.

[0040] The disclosed embodiments provide a system framework for implementing a personalized product recommendation method. The system obtains a user embedding vector representing user interests and a product embedding vector representing candidate products, maps the user embedding vector and the product embedding vector to obtain a user feedback vector, refines the user embedding vector and the user feedback vector to obtain a vector refined representation, and then calculates the user interest representation in the vector refined representation to obtain a related interest representation. A recommendation feature vector is generated based on the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation. The user click probability of the candidate product is predicted based on the recommendation feature vector, thereby determining the recommendation target. In this way, user interest information is transferred to candidate product information through inter-domain mapping, and the vector is refined by performing a self-attention operation to obtain a vector refined representation. The click probability of the candidate product is predicted based on the user interest in the vector refined representation. Compared to personalized recommendations for current purchases based on historical shopping data, this method avoids the noise impact on candidate products caused by sparse historical data or heterogeneous cross-domain data, thereby improving the fit between recommended products and user interests.

[0041] Optionally, obtaining the user embedding vector and the product embedding vector includes: obtaining the interest information of the target user and the product information of the candidate products; embedding the interest information and the product information into low-dimensional dense vectors respectively, to obtain the user embedding vector corresponding to the interest information and the product embedding vector corresponding to the product information.

[0042] In some embodiments, interest information and product information are densified through embedding technology to obtain low-dimensional dense vectors, aiming to use low-dimensional, dense vector space to represent high-dimensional, sparse vector space. The learned features can be used for machine learning tasks such as classification, regression, and clustering.

[0043] In some embodiments, a set of features (f1, f2, ..., f N ) is embedded into a low-dimensional dense vector e∈R N×d In the example above, we get the user embedding vector e user and product embedding vector e item , where N is the number of features of a set of features of interest information or product information, d is the embedding size of the low-dimensional dense vector, and the embedding size of the low-dimensional dense vector ranges from 8 to 16.

[0044] Optionally, after obtaining the target user's interest information and the product information of the candidate products, and before embedding the interest information and the product information into low-dimensional dense vectors respectively, the method also includes: performing data processing on the interest information and the product information, wherein the data processing includes one or more of data cleaning, data completion, and data standardization.

[0045] In some embodiments, since the user behavior data between the interest domain (interest information) and the product domain (product information) is highly heterogeneous, the user embedding vector and the product embedding vector obtained through network embedding technology have large differences in distribution. It is necessary to transfer the user interest in the interest domain to the product domain through inter-domain mapping. Therefore, it is necessary to map the user embedding vector and the product embedding vector to each other to obtain a user feedback vector.

[0046] Optionally, a user feedback vector is obtained by mapping the user embedding vector and the product embedding vector to each other, including: establishing an orthogonal mapping model, wherein the orthogonal mapping model includes a first mapping model and / or a second mapping model; iteratively updating the orthogonal mapping model based on a dual learning mechanism, and determining an iterative loss value; when the iterative loss value converges, determining that the iterative update of the orthogonal mapping model is completed; orthogonally mapping the user embedding vector to the product embedding vector through the first mapping model to obtain a first feedback vector, and orthogonally mapping the product embedding vector to the user embedding vector through the second mapping model to obtain a second feedback vector; and determining the first feedback vector and / or the second feedback vector as the user feedback vector.

[0047] In some embodiments, as Figure 3 As shown, the embodiment of the present disclosure provides a schematic diagram of the mapping between the embedding vector of domain A and the embedding vector of domain B, and the optimization process of learning the transition mapping is expressed by formula (1-1):

[0048]

[0049] Among them, e A is the embedding vector of domain A, D A is the vector set of embedding vectors of domain A, e B is the embedding vector of B domain, DB is the vector set of B domain embedding vectors, M AB represents the mapping model that maps the embedding vector of domain A to the embedding vector of domain B, M BA represents the mapping model that maps the B-domain embedding vector to the A-domain embedding vector, |||| is the Euclidean distance between the A-domain embedding vector and the B-domain embedding vector, and L is a part of the loss function, which represents the optimization of the minimum Euclidean distance between the A-domain embedding vector and the B-domain embedding vector.

[0050] In some embodiments, an orthogonal mapping model is established by formula (1-1), which can solve the problem of information transmission between domain A and domain B. However, there is still a negative transfer problem between domain A and domain B due to data sparsity. In order to solve the above problem and improve the information recommendation ability in domain A and domain B, it is necessary to introduce a dual learning mechanism, optimize the orthogonal mapping model in an iterative manner, and constrain the mapping on the two domains to a pair of orthogonal mappings. Among them, the use of orthogonal mapping can not only maintain the internal machine of different user embeddings, strengthen the structural invariance of user interests in each domain, but also prevent the performance degradation of a single domain. At the same time, in order to get rid of the differential interference between user embedding vectors and product embedding vectors, orthogonal mapping is performed separately when performing inter-domain mapping.

[0051] In some embodiments, the iterative process of learning the first mapping model and the second mapping model is identified based on formula (1-2):

[0052]

[0053] Among them, e user Embedding vector for user, D user is the vector set of user embedding vectors, e item is the product embedding vector, D item is a vector set of product embedding vectors, X represents the mapping model that orthogonally maps the user embedding vector to the product embedding vector, X T represents the mapping model that orthogonally maps the product embedding vector to the user embedding vector, |||| is the Euclidean distance between the user embedding vector and the product embedding vector, and L is a part of the loss function that represents the optimization of the minimum Euclidean distance between the user embedding vector and the product embedding vector.

[0054] In some embodiments, the first feedback vector Xe obtained by orthogonal mapping of the user embedding vector to the product embedding vector is obtained by orthogonal mapping model. item , the second feedback vector X where the product embedding vector is orthogonally mapped to the user embedding vector T e user .

[0055] In some embodiments, obtaining the second feedback vector X may be abandoned. T euser , the user click probability of the candidate product can be predicted through the first feedback vector.

[0056] Optionally, based on the self-attention mechanism, the user embedding vector and the user feedback vector are calculated respectively to obtain a vector refined representation: the first query matrix, the first key matrix and the first value matrix of the vector to be refined are calculated, wherein the vector to be refined includes the user embedding vector or the user feedback vector; calculation is performed according to the first query matrix, the first key matrix and the first value matrix to obtain the first attention head output corresponding to the vector to be refined; the first attention head outputs are spliced ​​to obtain a multi-head attention output; the multi-head attention outputs are sequentially subjected to feedforward network processing, residual link, degradation layer and normalization processing to obtain a vector refined representation corresponding to the vector to be refined.

[0057] In some embodiments, the user embedding vector and the product embedding vector are processed through a multi-head self-attention mechanism to obtain a more refined vector refined representation, where the vector refined representation includes a vector refined representation corresponding to the user embedding vector and a vector refined representation corresponding to the product embedding vector.

[0058] In some embodiments, the first query matrix, the first key matrix, and the first value matrix of the vector to be refined are calculated using formulas (1-3):

[0059]

[0060] Among them, Q1 is the first query matrix, K1 is the first key matrix, V1 is the first value matrix, e re is the vector to be refined, W1 Q 、W1 K 、W1 V are the linear matrices corresponding to the first query matrix, the first key matrix, and the first value matrix respectively.

[0061] In some embodiments, the user embedding vector e user , the first feedback vector Xe item , the second feedback vector X T e user One of the vectors is determined as the vector to be refined.

[0062] In some embodiments, the first attention head output corresponding to the vector to be refined is calculated using formula (1-4):

[0063]

[0064] Among them, head re (i) is the output of the first attention head corresponding to the vector to be refined, d re are the matrix dimensions of the first query matrix, the first key matrix, and the first value matrix.

[0065] In some embodiments, the first attention head output is concatenated using formula (1-5) to obtain the multi-head attention output:

[0066] multihead=Concat[head re (1),head re (2),...,head re (k)]W O Formula (1-5),

[0067] Among them, multihead is the multi-head attention output, k is the number of the first attention head output, W O ∈R d ' ×d ' is a linear matrix.

[0068] In some embodiments, the vector refined representation is obtained by formula (1-6):

[0069] R=FFN(multihead) formula (1-6),

[0070] Among them, R is the vector refined representation, FFN() is the feedforward network processing, residual link, degradation layer and normalization processing.

[0071] In some embodiments, the vector refined representation includes the user embedding vector e user The corresponding vector refined representation R1, the first feedback vector Xe item The corresponding vector refined representation R2, the second feedback vector X T e user The corresponding vector refinement represents a part of R3.

[0072] In some embodiments, due to the problems of overfitting caused by the increase of model capacity, the vanishing gradient causing the training process to become unstable, and models with more parameters often requiring more training time, the above problems can be improved through residual connections, layer normalization and degradation layers.

[0073] Optionally, extracting user interest representation from each vector refined representation includes: calculating a second query matrix, a second key matrix, and a second value matrix of the vector refined representation; performing calculations based on the second query matrix, the second key matrix, and the second value matrix to obtain a second attention head output; and using the second attention head output as the user interest representation of the vector refined representation.

[0074] In some embodiments, the second query matrix, the second key matrix, and the second value matrix of the vector refined representation are calculated by formulas (1-7):

[0075]

[0076] Among them, Q2 is the second query matrix, K2 is the second key matrix, V2 is the second value matrix, e int is a vector refined representation, are the linear matrices corresponding to the second query matrix, the second key matrix and the second value matrix respectively.

[0077] In some embodiments, the vector refined representation includes

[0078] In some embodiments, the second attention head output is obtained by calculating the second query matrix, the second key matrix, and the second value matrix using formula (1-8):

[0079]

[0080] Among them, head int (j) is the vector refined representation corresponding to the j-th second attention head output, d int are the matrix dimensions of the second query matrix, the second key matrix, and the second value matrix.

[0081] In some embodiments, the user interest vector is expressed by formula (1-9):

[0082] H={head int (1),head int (2),...,head int (l)} Formula (1-9),

[0083] Among them, H is the user interest vector, which reflects the different interests of users in the vector refined representation, and J is the number of user interest representations (i.e., the output of the second attention head).

[0084] In some embodiments, the user interest vectors include a user interest vector H1 corresponding to the vector refined representation R1, a user interest vector H2 corresponding to the vector refined representation R2, and a user interest vector H3 corresponding to the vector refined representation R3.

[0085] Optionally, calculations are performed based on the extracted user interest representations to obtain relevant interest representations corresponding to each vector refined representation, including: calculations are performed based on the product embedding vector and the user interest representation to obtain a first weight corresponding to each user interest representation; and weighted calculations are performed on the user interest representations based on the first weight to obtain relevant interest representations corresponding to the vector refined representation.

[0086] In some embodiments, the correlation between user interests and candidate products after forward mapping is different. By introducing a self-attention mechanism, each user interest representation is extracted, and a first weight corresponding to each user interest representation is set, so that the user interest representation is weighted calculated according to the first weight to obtain the relevant interest representation corresponding to the vector refined representation.

[0087] In some embodiments, the first weight corresponding to the user interest representation is determined by formula (1-10):

[0088]

[0089] Among them, ω j is the first weight corresponding to the jth user interest representation, head int (j) is the jth user interest representation corresponding to the vector refined representation, e item is the product embedding vector, W3 is the preset linear matrix, and J is the number of user interest representations.

[0090] In some embodiments, the relevant interest representation corresponding to the vector refined representation is determined by formula (1-11):

[0091]

[0092] Among them, I is the relevant interest representation corresponding to the vector refined representation, ω j is the first weight corresponding to the jth user interest representation, head int (j) is the j-th user interest representation corresponding to the vector refined representation.

[0093] In some embodiments, the related interests corresponding to the vector refined representation R1 are represented as I1, the related interests corresponding to the vector refined representation R2 are represented as I2, and the related interests corresponding to the vector refined representation R3 are represented as I3.

[0094] In some embodiments, the relevant interest representation corresponding to the vector refined representation is the highest interest representation in the vector refined representation.

[0095] Optionally, a recommendation feature vector is generated by at least one of the following methods: concatenating the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation to obtain a recommendation feature vector; respectively calculating the second weights of the vector refined representation and the related interest representation, and fusing the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation according to the second weights to obtain a recommendation feature vector.

[0096] In some embodiments, the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation are spliced ​​to obtain the recommendation feature vector, including: embedding the user vector e user, product embedding vector e item , the vector refined representation R1 corresponding to the user embedding vector, the vector refined representation R2 corresponding to the first feedback vector, the related interest representation I1 corresponding to the user embedding vector, and the related interest representation I2 corresponding to the first feedback vector are spliced ​​to obtain the recommended feature vector.

[0097] In some embodiments, the method of obtaining the recommendation feature vector by direct concatenation ignores the different importance of different types of interests. Therefore, the second weight is used for fusion to consider the importance of different types of interests, thereby improving the prediction accuracy.

[0098] In some embodiments, the second weights of the calculated vector refined representation and the related interest representation are determined by formula (1-12):

[0099]

[0100] Among them, ω * is the second weight corresponding to R1, R2, I1, and I2, W * is the preset matrix corresponding to R1, R2, I1, I2, f * T are the trainable preset vector parameters corresponding to R1, R2, I1, and I2, t * are preset scalars corresponding to R1, R2, I1, and I2, and their purpose is to ensure that all types of interest representations participate in the final recommendation prediction.

[0101] In some embodiments, the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation are fused according to the second weight using formula (1-13) to obtain a recommendation feature vector:

[0102]

[0103] Among them, C is the recommended feature vector.

[0104] In some embodiments, the probabilistic prediction model is a multi-layer perceptron (MLP).

[0105] Combine Figure 4As shown, the embodiment of the present disclosure provides a personalized product recommendation system, including an acquisition module 401, a mapping module 402, an extraction module 403 and a determination module 404; the acquisition module 401 is used to obtain a user embedding vector and a product embedding vector, wherein the user embedding vector is used to represent the interest information of the target user, and the product embedding vector is used to represent the product information of the candidate product; the mapping module 402 is used to map the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and calculate the user embedding vector and the user feedback vector respectively based on the self-attention mechanism to obtain a vector refined representation; the extraction module 403 is used to extract the user interest representation from each vector refined representation, and calculate based on the extracted user interest representation to obtain the relevant interest representation corresponding to each vector refined representation; the determination module 404 is used to calculate the recommendation feature vector according to a preset probability prediction model to obtain the user click probability of the candidate product, and determine the recommendation target from the candidate product according to the user click probability, wherein the recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the relevant interest representation.

[0106] The disclosed embodiments provide a method for personalized product recommendation. This method obtains a user embedding vector representing user interests and a product embedding vector representing candidate products, maps the user embedding vector and the product embedding vector to obtain a user feedback vector, refines the user embedding vector and the user feedback vector to obtain a vector refined representation, and then calculates the user interest representation in the vector refined representation to obtain a related interest representation. A recommendation feature vector is generated based on the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation. The user click probability of the candidate product is predicted based on the recommendation feature vector, thereby determining the recommendation target. In this way, user interest information is transferred to candidate product information through inter-domain mapping, and the vector is refined by performing a self-attention operation to obtain a vector refined representation. The click probability of the candidate product is predicted based on the user interest in the vector refined representation. Compared to personalized recommendations for current purchases based on historical shopping data, this method avoids the noise impact on candidate products caused by sparse historical data or heterogeneous cross-domain data, thereby improving the fit between recommended products and user interests.

[0107] Figure 5 The following is a schematic diagram showing the structure of a computer system suitable for implementing an electronic device according to an embodiment of the present application. Figure 5 The computer system 500 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0108] like Figure 5As shown, the computer system 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 502 or the program loaded from the storage part 508 into the random access memory (RAM) 503, such as executing the method in the above embodiment. Various programs and data required for system operation are also stored in the RAM 503. The CPU 501, ROM 502 and RAM 503 are connected to each other via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0109] The following components are connected to the I / O interface 505: an input section 506 including a keyboard, a mouse, and the like; an output section 507 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 508 including a hard disk; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I / O interface 505 as needed. Removable media 511, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 510 as needed, so that computer programs read therefrom can be installed into the storage section 508 as needed.

[0110] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 509, and / or installed from a removable medium 511. When the computer program is executed by the central processing unit (CPU) 501, the various functions defined in the system of the present application are executed.

[0111] It should be noted that the computer-readable medium shown in the embodiments of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, which carries a computer-readable computer program. This propagated data signal can take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. A computer program embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0112] The embodiments of the present disclosure further provide a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, any one of the methods in the embodiments is implemented.

[0113] Regarding the computer-readable storage media in the embodiments of the present disclosure, those skilled in the art will understand that all or part of the steps in implementing the aforementioned method embodiments can be accomplished by hardware associated with the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the aforementioned method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0114] The electronic device disclosed in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication with each other. The memory is used to store computer programs, the communication interface is used to communicate, and the processor and the transceiver are used to run the computer program, so that the electronic device executes each step of the above method.

[0115] In this embodiment, the memory may include a random access memory (RAM), and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage.

[0116] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a graphics processing unit (GPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0117] The above description and the accompanying drawings fully illustrate the embodiments of the present disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, process and other changes. The embodiments represent only possible variations. Unless expressly required, individual components and functions are optional, and the order of operations may vary. Parts and subsamples of some embodiments may be included in or replace parts and subsamples of other embodiments. Moreover, the terms used in this application are only used to describe the embodiments and are not used to limit the claims. As used in the description of the embodiments and claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to also include the plural forms. Similarly, the term "and / or" as used in this application refers to any and all possible combinations of one or more associated listings. In addition, when used in this application, the term "comprise" and its variations "comprises" and / or comprising refer to the presence of a stated subsample, whole, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other subsamples, wholes, steps, operations, elements, components and / or groups of these. In the absence of further restrictions, an element defined by the statement "comprises a..." does not exclude the presence of other identical elements in the process, method or device that includes the element. In this article, each embodiment may focus on the differences from other embodiments, and the same and similar parts between the various embodiments can be referenced to each other. For the methods, products, etc. disclosed in the embodiments, if they correspond to the method part disclosed in the embodiments, then the relevant parts can be found in the description of the method part.

[0118] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software may depend on the specific application and design constraints of the technical solution. Technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present disclosure. Technicians can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0119] In the embodiments disclosed herein, the disclosed methods and products (including but not limited to devices, equipment, etc.) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units can be merely a logical functional division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some sub-samples can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to implement this embodiment. In addition, the functional units in the embodiments of the present disclosure can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0120] The flowcharts and block diagrams in the accompanying drawings show the possible implementation architectures, functions and operations of the systems, methods and computer program products according to the embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of the code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical functions. In some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, which can depend on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different boxes can also occur in an order different from that disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, which can depend on the functions involved. Each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs the specified function or action, or may be implemented by a combination of dedicated hardware and computer instructions.

Claims

1. A personalized product recommendation method, characterized in that: include: Obtaining a user embedding vector and a product embedding vector, wherein the user embedding vector is used to represent the interest information of the target user, and the product embedding vector is used to represent the product information of the candidate product; A user feedback vector is obtained by mapping the user embedding vector and the product embedding vector to each other: an orthogonal mapping model is established, wherein the orthogonal mapping model includes a first mapping model and / or a second mapping model; the orthogonal mapping model is iteratively updated in a dual learning mechanism, and an iterative loss value is determined; when the iterative loss value converges, it is determined that the iterative update of the orthogonal mapping model is completed; the user embedding vector is orthogonally mapped to the product embedding vector through the first mapping model to obtain a first feedback vector, and the product embedding vector is orthogonally mapped to the user embedding vector through the second mapping model to obtain a second feedback vector; the first feedback vector and / or the second feedback vector are determined as the user feedback vector; Based on the self-attention mechanism, the user embedding vector and the user feedback vector are respectively calculated to obtain a vector refined representation: a first query matrix, a first key matrix, and a first value matrix of the vector to be refined are calculated, wherein the vector to be refined includes the user embedding vector or the user feedback vector; a first attention head output corresponding to the vector to be refined is obtained by calculating according to the first query matrix, the first key matrix, and the first value matrix; the first attention head outputs are spliced ​​to obtain a multi-head attention output; the multi-head attention output is sequentially subjected to feedforward network processing, residual link, degradation layer, and normalization processing to obtain a vector refined representation corresponding to the vector to be refined; Extracting user interest representations from each of the vector refined representations, and performing calculations based on the extracted user interest representations to obtain relevant interest representations corresponding to each of the vector refined representations: performing calculations based on the product embedding vectors and the user interest representations to obtain first weights corresponding to each of the user interest representations; performing weighted calculations on the user interest representations based on the first weights to obtain relevant interest representations corresponding to the vector refined representations; The recommendation feature vector is calculated according to a preset probability prediction model to obtain the user click probability of the candidate product, and the recommendation target is determined from the candidate products based on the user click probability, wherein the recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation.

2. The method according to claim 1, characterized in that Obtain user embedding vectors and product embedding vectors, including: Obtain target user interest information and candidate product information; The interest information and the product information are respectively embedded in low-dimensional dense vectors to obtain a user embedding vector corresponding to the interest information and a product embedding vector corresponding to the product information.

3. The method according to claim 2, characterized in that After obtaining the target user's interest information and the product information of the candidate products, and before embedding the interest information and the product information into low-dimensional dense vectors, the method further includes: Data processing is performed on the interest information and the product information, wherein the data processing includes one or more of data cleaning, data completion, and data standardization.

4. The method according to claim 1, wherein Extracting user interest representations from each of the vector refined representations includes: calculating a second query matrix, a second key matrix, and a second value matrix of the vector refined representation; Performing calculations based on the second query matrix, the second key matrix, and the second value matrix to obtain a second attention head output; The second attention head outputs the user interest representation as the vector refined representation.

5. The method according to any one of claims 1 to 4, characterized in that Generate a recommendation feature vector using at least one of the following methods: Concatenate the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation to obtain a recommendation feature vector; Second weights of the vector refined representation and the related interest representation are calculated respectively, and the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation are fused according to the second weights to obtain a recommendation feature vector.

6. A personalized product recommendation system, characterized in that: include: An acquisition module is used to acquire a user embedding vector and a product embedding vector, wherein the user embedding vector is used to represent the interest information of the target user, and the product embedding vector is used to represent the product information of the candidate product; a mapping module, configured to map the user embedding vector and the product embedding vector to each other to obtain a user feedback vector, and to calculate the user embedding vector and the user feedback vector based on a self-attention mechanism to obtain a vector refined representation; The mapping module obtains a user feedback vector by: establishing an orthogonal mapping model, wherein the orthogonal mapping model includes a first mapping model and / or a second mapping model; iteratively updating the orthogonal mapping model using a dual learning mechanism and determining an iterative loss value; determining that the iterative update of the orthogonal mapping model is complete when the iterative loss value converges; orthogonally mapping the user embedding vector to the product embedding vector using the first mapping model to obtain a first feedback vector, and orthogonally mapping the product embedding vector to the user embedding vector using the second mapping model to obtain a second feedback vector; and determining the first feedback vector and / or the second feedback vector as the user feedback vector; The mapping module obtains a vector refined representation by: calculating a first query matrix, a first key matrix, and a first value matrix of a vector to be refined, wherein the vector to be refined includes the user embedding vector or the user feedback vector; performing calculations based on the first query matrix, the first key matrix, and the first value matrix to obtain a first attention head output corresponding to the vector to be refined; concatenating the first attention head outputs to obtain a multi-head attention output; and sequentially performing feedforward network processing, residual linking, degradation layer processing, and normalization processing on the multi-head attention output to obtain a vector refined representation corresponding to the vector to be refined; an extraction module, configured to extract user interest representations from each of the vector refined representations, and perform calculations based on the extracted user interest representations to obtain relevant interest representations corresponding to each of the vector refined representations; The extraction module obtains the relevant interest representation by: performing calculations based on the product embedding vector and the user interest representation to obtain a first weight corresponding to each of the user interest representations; performing weighted calculations on the user interest representations based on the first weights to obtain the relevant interest representation corresponding to the vector refined representation; A determination module is used to calculate the recommendation feature vector according to a preset probability prediction model to obtain the user click probability of the candidate product, and determine the recommendation target from the candidate products based on the user click probability, wherein the recommendation feature vector is generated by the user embedding vector, the product embedding vector, the vector refined representation, and the related interest representation.

7. An electronic device, characterized in that: include: processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images