Commodity recommendation method and system and storage medium

By building an initial recommendation model, analyzing user data and trust dissemination, and obtaining user and product characteristics, the problem of unconsidered trust dissemination in the existing technology is solved, and the accuracy of product recommendation is improved.

CN118964737APending Publication Date: 2024-11-15GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411039982.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing product recommendation methods fail to effectively consider the impact of trust dissemination, resulting in low recommendation accuracy.

Method used

Build an initial recommendation model, including reliability analysis unit, modeling unit and scoring prediction unit, perform reliability analysis and modeling through user data, obtain feature representation vectors of users and products, and optimize the recommendation model through loss function to capture trust propagation Influence and improve recommendation accuracy.

Benefits of technology

By obtaining user reliability values ​​from user data, analyzing user and product characteristics, improving the accuracy and effect of product recommendations, making the recommendation results more accurate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964737A_ABST
    Figure CN118964737A_ABST
Patent Text Reader

Abstract

The invention provides a commodity recommendation method and system and a storage medium, and relates to the technical field of data recommendation. The method comprises the following steps: importing user data collected from a shopping platform to construct an initial recommendation model, carrying out reliability analysis on the user data to obtain a reliability value, carrying out user modeling on the user data and the reliability value to obtain a user feature representation vector, carrying out commodity modeling on the user data and the reliability value to obtain a commodity feature representation vector, and carrying out recommendation on the commodity feature representation vector. Performing fusion calculation on the user feature representation vector and the commodity feature representation vector to obtain a prediction score; iteratively optimizing the initial recommendation model based on the prediction score through a loss function to obtain a recommendation model; and performing commodity recommendation for the user through the recommendation model to obtain a commodity recommendation result. The commodity score is analyzed by using the reliability value, the capture of trust propagation is completed, and the potential features of the user and the commodity are learned to mine the preference and interest of the user to the commodity, so that the recommendation result is more accurate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention mainly relates to the technical field of data recommendation, and in particular to a commodity recommendation method, system and storage medium. Background Art

[0002] With the rapid development of the Internet, network information has increased dramatically. The problem faced by product recommendations has changed from the lack of information in the past to how to effectively filter information and obtain valuable information in the current information overload environment. Recommendation methods are a type of information filtering technology that aims to predict user preferences and recommend products or content that may be of interest to users based on them. They are widely used in e-commerce, streaming services, personalized learning and other fields.

[0003] In real life, a user's preference for a product is not only influenced by the product preferences of their social friends, but also to a certain extent by the product preferences of their friends' friends, which is called trust propagation. However, existing recommendation methods do not consider the influence of trust propagation, resulting in low recommendation accuracy. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a product recommendation method, system and storage medium in view of the deficiencies in the prior art.

[0005] The technical solution of the present invention to solve the above technical problems is as follows:

[0006] A product recommendation method comprises the following steps:

[0007] Constructing an initial recommendation model, and importing user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit;

[0008] Performing reliability analysis on the user data by the reliability analysis unit of the initial recommendation model to obtain a reliability value;

[0009] Performing user modeling on the user data and the reliability value through the modeling unit of the initial recommendation model to obtain a user feature representation vector, and performing product modeling on the user data and the reliability value to obtain a product feature representation vector;

[0010] The score prediction unit of the initial recommendation model performs a fusion calculation on the user feature representation vector and the product feature representation vector to obtain a predicted score;

[0011] The predicted score and the user data are calculated by a loss function to obtain a loss value, and the initial recommendation model is optimized according to the loss value to obtain a recommendation model;

[0012] The recommendation model is used to calculate the imported user nodes to be recommended and the commodity nodes to be recommended to obtain commodity recommendation results.

[0013] Another technical solution of the present invention to solve the above technical problems is as follows:

[0014] A product recommendation system, comprising:

[0015] An import module is used to construct an initial recommendation model and import user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit;

[0016] An analysis module, configured to perform reliability analysis on the user data through a reliability analysis unit of the initial recommendation model to obtain a reliability value;

[0017] A modeling module, configured to perform user modeling on the user data and the reliability value through a modeling unit of the initial recommendation model to obtain a user feature representation vector, and perform commodity modeling on the user data and the reliability value to obtain a commodity feature representation vector;

[0018] A prediction module, configured to perform a fusion calculation on the user feature representation vector and the product feature representation vector through a score prediction unit of the initial recommendation model to obtain a predicted score;

[0019] An optimization module, configured to calculate the predicted score and the user data through a loss function to obtain a loss value, and optimize the initial recommendation model according to the loss value to obtain a recommendation model;

[0020] The recommendation module is used to calculate the imported to-be-recommended user nodes and to-be-recommended commodity nodes through the recommendation model to obtain commodity recommendation results.

[0021] Another technical solution of the present invention to solve the above technical problems is as follows:

[0022] A product recommendation system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a product recommendation method as described above is implemented.

[0023] Another technical solution of the present invention to solve the above technical problems is as follows:

[0024] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the above-mentioned commodity recommendation method is implemented.

[0025] The beneficial effects of the present invention are as follows: by obtaining user reliability values ​​from user data, the user reliability values ​​are used to effectively analyze user product ratings during user modeling and product modeling, so as to capture trust propagation, learn the characteristic representations of users and products, explore users' preferences and interests in products, improve the effect of product recommendations, and make the recommendation results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flowchart of a product recommendation method provided by an embodiment of the present invention;

[0027] Figure 2 A social network diagram of user nodes provided by an embodiment of the present invention;

[0028] Figure 3 This is a module block diagram of a product recommendation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0030] like Figure 1 As shown, a commodity recommendation method provided by an embodiment of the present invention includes the following steps:

[0031] Constructing an initial recommendation model, and importing user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit;

[0032] Performing reliability analysis on the user data by the reliability analysis unit of the initial recommendation model to obtain a reliability value;

[0033] Performing user modeling on the user data and the reliability value through the modeling unit of the initial recommendation model to obtain a user feature representation vector, and performing product modeling on the user data and the reliability value to obtain a product feature representation vector;

[0034] The score prediction unit of the initial recommendation model performs a fusion calculation on the user feature representation vector and the product feature representation vector to obtain a predicted score;

[0035] The predicted score and the user data are calculated by a loss function to obtain a loss value, and the initial recommendation model is optimized according to the loss value to obtain a recommendation model;

[0036] The recommendation model is used to calculate the imported user nodes to be recommended and the commodity nodes to be recommended to obtain commodity recommendation results.

[0037] Specifically, in the initial recommendation model, the output of the reliability analysis unit is connected to the input of the modeling unit, and the output of the modeling unit is connected to the input of the rating prediction unit.

[0038] It should be understood that multiple user nodes to be recommended and multiple product nodes to be recommended are imported into the recommendation model, and recommendation predictions are made for the multiple user nodes to be recommended and multiple product nodes to be recommended through the recommendation model to obtain predicted scores of each user node to be recommended for the multiple product nodes to be recommended, and the predicted scores are sorted in descending order, and the predicted scores within the preset range of the sorting results are selected, and the products corresponding to the selected predicted scores are recommended to the corresponding users (i.e., the product recommendation results).

[0039] In the embodiment of the present invention, the user reliability value is obtained from the user data, and the user reliability value is used to scale the user product rating (i.e., strengthen or weaken the recommendation) during the user modeling and product modeling process to complete the capture of trust propagation. The user feature representation vector contains the user's initial preference and the user's social information, and the product feature representation vector contains the popularity of the product. By learning the feature representation of users and products, the user's preference and interest in the product can be explored, the effect of product recommendation can be improved, and the recommendation result can be made more accurate.

[0040] Preferably, the user data includes a plurality of user nodes and a plurality of neighbor sets corresponding to the user nodes respectively;

[0041] The reliability analysis unit of the initial recommendation model performs reliability analysis on the user data to obtain a reliability value, specifically:

[0042] Counting the number of neighbors of the first-order neighbors of the corresponding user node according to the plurality of neighbor sets to obtain a sequence of the number of neighbors of the plurality of user nodes;

[0043] The plurality of neighbor quantity sequences are respectively converted by encoding using a reliability expression to obtain a context vector corresponding to each of the user nodes, and reliability calculations are performed on the plurality of context vectors to obtain a reliability value corresponding to each of the user nodes. The reliability expression is:

[0044]

[0045] Among them, c i is user node u i The context vector of , Encoder is a sequence encoder composed of LSTM units, is user node u i The number of neighbors of the first-order neighbors of user node u i Neighbor Set is user node ui The first-order neighbor of The number of neighbors, r i is user node u i The reliability value of minC is the minimum value of the context vector C of all user nodes, and maxC is the maximum value of the context vector C of all user nodes.

[0046] It should be understood that the user data includes multiple user nodes and neighbor sets corresponding to the multiple user nodes, and the total number of neighbor nodes in the neighbor set corresponding to each user node is counted. Figure 2 As shown in the figure, the total number of neighbor nodes of user node u1 is 5, the total number of neighbor nodes of user node u2 is 3, the total number of neighbor nodes of user node u3 is 3, the total number of neighbor nodes of user node u4 is 3, the total number of neighbor nodes of user node u5 is 4, and the total number of neighbor nodes of user node u6 is 1. The first-order neighbors of user node u1 include u2, u3, u4, u5, and u6, and the sequence of the number of neighbors of the first-order neighbors of user node u1 is <3,3,3,4,1>.

[0047] Among them, the context vector in the reliability expression is the norm corresponding to the context vector, and the Euclidean norm corresponding to the context vector of all user nodes is obtained by the numpy.linalg.norm() function, and all Euclidean norms are compared to obtain the maximum value maxC and the minimum value minC.

[0048] In an embodiment of the present invention, by quantifying the number of neighbors of the neighbors of a user in a social network (i.e., a shopping platform), the corresponding neighbor number sequence is converted into a context vector of the user node. This vector contains the global information and key features of the neighbor number sequence, and then the reliability value corresponding to the user is obtained. The reliability value can measure the trust and influence of the user in the social network. The size of the reliability value indicates the strength of the user's social ability. The larger the value, the stronger the user's social ability, and the more reference significance the user's score has, thereby improving the accuracy and credibility of the recommendation results.

[0049] Preferably, the user data further includes a plurality of commodity nodes and commodity ratings corresponding to the plurality of user nodes;

[0050] The modeling unit of the initial recommendation model performs user modeling on the user data and the reliability value to obtain a user feature representation vector, specifically:

[0051] Vectorizing the plurality of user nodes, the neighbor sets corresponding to the plurality of user nodes, the plurality of product nodes, and the product ratings corresponding to the plurality of user nodes, respectively, to obtain a plurality of user embedding vectors, a neighbor embedding vector corresponding to each of the user nodes, a plurality of product embedding vectors, and a rating embedding vector corresponding to each of the user nodes;

[0052] Performing product aggregation on the plurality of user embedding vectors, the plurality of product embedding vectors, the plurality of reliability values, and the plurality of rating embedding vectors respectively to obtain a user-product embedding vector corresponding to each of the user nodes, wherein the user-product embedding vector corresponding to each of the user nodes includes a neighbor-product embedding vector corresponding to each neighbor node in each neighbor set;

[0053] Performing social aggregation on the plurality of user embedding vectors and the plurality of neighbor product embedding vectors respectively to obtain a user social embedding vector corresponding to each user node;

[0054] The user feature expression is used to calculate each user-item embedding vector and the corresponding user-social embedding vector, respectively, to obtain the user feature representation vector corresponding to each user node. The user feature expression is:

[0055]

[0056] Among them, h i is user node u i The user feature representation vector is is user node u i The user-item embedding vector of is user node u i The user social embedding vector, W i is the weight matrix.

[0057] In the embodiment of the present invention, the user-product embedding vector and the user-social embedding vector are concatenated and then dimensionally transformed to obtain a synthetic vector, which can represent the user's preferences and the user's social information.

[0058] Preferably, the performing of product aggregation on the plurality of user embedding vectors, the plurality of product embedding vectors, the plurality of reliability values ​​and the plurality of rating embedding vectors to obtain the user-product embedding vector corresponding to each user node is specifically as follows:

[0059] The plurality of reliability values, the plurality of product embedding vectors and the plurality of rating embedding vectors are calculated respectively through an interaction vector expression to obtain an interaction representation vector of each user node for the rated product. The interaction vector expression is:

[0060]

[0061] Among them, x ij is user node u i For product node v j The interaction representation vector, f U is a multi-layer perceptron MLP, q j is the commodity node v j The corresponding product embedding vector, e ij is user node u i For product node v j The rating embedding vector corresponding to the rating, r i is user node u i The reliability value of It is the concatenation operation of two vectors;

[0062] Attention analysis is performed on the multiple user embedding vectors and the corresponding interaction representation vectors through the product embedding expression to obtain the attention corresponding to the multiple interaction representation vectors, and the multiple attentions are allocated to the corresponding interaction representation vectors. The attention allocation results are aggregated to obtain the user product embedding vector corresponding to each user node. The product embedding expression is:

[0063]

[0064] Among them, α ij is user node u i For product node v j The attention level, p i is user node u i The corresponding user embedding vector, is user node u i The product embedding vector, x ij is user node u i For product node v j The interaction representation vector is is user node u i The set of evaluated products, ReLU is the activation function, Tanh is the activation function, W I is the weight matrix, b I is the bias vector, T is the matrix transpose, It is the concatenation operation of two vectors.

[0065] Specifically, the user nodes corresponding to the product embedding vector and the rating embedding vector are derived through the multi-layer perceptron MLP, and the reliability value corresponding to the user node is assigned to the rating embedding vector to obtain the user's interaction representation vector for the rated product; that is, the reliability value is used to measure the reference value of user evaluation in social communication, so the reliability value is assigned to the rating embedding vector.

[0066] Through a two-layer neural network, multiple user embedding vectors and corresponding interaction representation vectors are learned (i.e., attention analysis), the attention of multiple corresponding interaction representation vectors is obtained, multiple attentions are assigned to corresponding interaction representation vectors, and multiple interaction representation vectors with attention are aggregated to obtain user-item embedding vectors corresponding to each user node; that is, when a user rates multiple items, the weight value of the interaction representation vector is assigned according to the attention, and after aggregation, the model focuses on the user's high-rated items to obtain the user's initial preference for the items. Among them, the two-layer neural network contains a neural network architecture with two hidden layers.

[0067] It should be understood that when each neighbor node in the neighbor set is a user node, the user-item embedding vector corresponding to the user node is the neighbor-item embedding vector corresponding to the neighbor node, that is, the user node u in the product embedding expression i Replaced with user node u k (i.e. user node u i Neighbor node u k ), the replaced product embedding expression is:

[0068]

[0069] Among them, α kj is user node u k For product node v j The attention level, p k is user node u k The corresponding user embedding vector, is user node u k The product embedding vector, x kj is user node u k For product node v j The interaction representation vector is is user node u k The set of evaluated products, ReLU is the activation function, Tanh is the activation function, W I is the weight matrix, b I is the bias vector, T is the matrix transpose, It is the concatenation operation of two vectors;

[0070] Calculating the neighbor item embedding vector When the attention involved is α kj middle, At this time, the commodity node v j Represents user node u k Any product that has been interacted with, and the user node u i Any commodity node v that has been interacted with j The symbols are the same, but the commodity nodes they represent are different.

[0071] In the embodiment of the present invention, the strength of the user's social ability is associated with the user's rating of the product, the user's connection and rating information for the product is captured, and the degree of association of the same user with different products is reflected through the attention degree, and the user's initial preference for the product is reflected through product aggregation.

[0072] Preferably, performing social aggregation on the plurality of user embedding vectors and the plurality of neighbor product embedding vectors respectively to obtain the user social embedding vector corresponding to each user node is specifically as follows:

[0073] Attention analysis is performed on the multiple user embedding vectors and the multiple neighbor product embedding vectors through the social embedding expression to obtain the user social strength corresponding to each user node, and the multiple user social strengths are assigned to the corresponding neighbor product embedding vectors, and the social strength assignment results are aggregated to obtain the user social embedding vector corresponding to each user node. The social embedding expression is:

[0074]

[0075] Among them, β ik is user node u i With neighbor node u k The social intensity, is user node u i Corresponding neighbor node u k The neighbor product embedding vector, p i is user node u i The user embedding vector of is user node u i The social embedding vector of is user node u i Neighbor set, ReLU is the activation function, Tanh is the activation function, , W s is the weight matrix, b s is the bias vector, is the concatenation operation of two vectors, and T is the matrix transpose.

[0076] Specifically, multiple user embedding vectors and multiple neighbor product embedding vectors are learned (attention analysis) through a two-layer neural network to obtain the user social strength; wherein the two-layer neural network includes a neural network architecture with two hidden layers.

[0077] In the embodiment of the present invention, the user's social strength is used to reflect the degree of association between the user and each neighbor node. Based on the user's social strength, the model is helped to pay attention to the product preferences of social friends with strong social relationships. Then, through social aggregation, the product interaction information of all the user's social friends is obtained, that is, the user's social embedding vector.

[0078] Preferably, the performing of commodity modeling on the user data and the reliability value to obtain a commodity feature representation vector is specifically as follows:

[0079] The plurality of reliability values, the plurality of user embedding vectors and the plurality of rating embedding vectors are calculated respectively through the interactive perception vector expression to obtain the interactive perception representation vector of the opinions of each user node on the rated product. The interactive perception vector expression is:

[0080]

[0081] Among them, y ji is the commodity node v j With user node u i The opinion interaction perception representation vector, f V is a multi-layer perceptron MLP, p i is user node u i The corresponding user embedding vector, e ij is user node u i For product node v j The rating embedding vector corresponding to the rating, r i is user node u i The reliability value of It is the concatenation operation of two vectors;

[0082] Attention analysis is performed on the multiple product embedding vectors and the corresponding opinion interaction perception representation vectors through the product feature expression to obtain the attention weights of the multiple opinion interaction perception representation vectors, and the multiple attention weights are assigned to the corresponding opinion interaction perception representation vectors. The attention weight assignment results are aggregated to obtain the product feature representation vector corresponding to each product node. The product feature expression is:

[0083]

[0084] Among them, μ ji is the attention weight, qj is the commodity node v j The corresponding product embedding vector, y ji is the commodity node v j With user node u i The opinion interaction perception representation vector, h j is the commodity node v j The feature representation vector of For the node v that has evaluated the product j The user set, ReLU is the activation function, Tanh is the activation function, W j is the weight matrix, b j is the bias vector, T is the matrix transpose, It is the concatenation operation of two vectors.

[0085] Specifically, a two-layer neural network is used to learn multiple product embedding vectors and corresponding opinion interaction perception representation vectors (i.e., attention analysis) to obtain attention weights. After the attention weights are assigned to the opinion interaction perception representation vectors, the opinion interaction perception representation vectors of different user nodes for the same product node are aggregated to obtain the feature representation vector of the product. The two-layer neural network includes a neural network architecture with two hidden layers.

[0086] In the embodiment of the present invention, the connection and rating information of the product to the user is captured according to the opinion interaction perception representation vector, and the degree of association of different users with the same product is reflected through the attention weight, and then the popularity of the product is obtained through user aggregation, that is, the product feature representation vector.

[0087] Preferably, the score prediction unit of the initial recommendation model performs a fusion calculation on the user feature representation vector and the product feature representation vector to obtain a predicted score, specifically:

[0088] The user feature representation vector corresponding to each user node and the commodity feature representation vector corresponding to each commodity node are calculated respectively through the score calculation expression to obtain the predicted score of each user node for each commodity node. The score calculation expression is:

[0089]

[0090] in, is user node u i For product node v j The prediction score, h i is user node u i The user feature representation vector, h j is the commodity node v jThe product feature representation vector, W1, W2, W r is the weight matrix, b1 and b2 are bias vectors, Tanh is the activation function, T is the matrix transpose, This is a vector concatenation operation.

[0091] In the embodiment of the present invention, the user's rating of each product is analyzed according to the popularity of the product, the user's preference for the product and the user's social relationship to determine whether the recommendation ability of the model is qualified.

[0092] Preferably, the predicted score and the user data are calculated by a loss function to obtain a loss value, and the initial recommendation model is optimized according to the loss value to obtain a recommendation model, specifically:

[0093] The predicted score and the product score are calculated by using an initial loss function to obtain an initial loss value. The initial loss function is:

[0094]

[0095] Among them, L is the initial loss value, r ij is user node u i For product node v j Product ratings, is user node u i For product node v j 's prediction score;

[0096] The initial recommendation model is iteratively optimized according to the initial loss value through the RMSprop algorithm to obtain a target loss function, and the model parameters of the initial recommendation model are updated according to the target loss function to obtain a recommendation model.

[0097] It should be understood that the RMSprop algorithm is an adaptive learning rate optimization algorithm, which is mainly used to solve the learning rate adjustment problem in gradient descent, such as slow convergence or non-convergence caused by improper learning rate selection.

[0098] In the embodiment of the present invention, the loss is calculated based on the predicted score and the actual score, and then the initial model is optimized based on the loss value to improve the recommendation accuracy of the model.

[0099] It should be understood that each weight matrix and each bias vector in this solution can be trained as parameters, and the training is:

[0100] First, the weight matrix and bias vector are randomly initialized using a function, and the weight matrix and bias vector are continuously updated through the back propagation algorithm and the RMSprop algorithm based on the initial loss value to minimize the initial loss function.

[0101] Specifically, the initialization function is used to initialize the weight matrix and bias vector, where the bias vector is initialized to zero; for the ReLU activation function, the Kaiming uniform distribution is used to initialize the weight matrix, and for the Tanh activation function, the Xavier uniform distribution is used to initialize the weight matrix; the gradient of the initial loss value with respect to the weight matrix and the bias vector is calculated through the back propagation algorithm, and each weight matrix and each bias vector is updated according to the calculated gradient through the RMSprop algorithm.

[0102] like Figure 3 As shown, a commodity recommendation system provided by an embodiment of the present invention includes:

[0103] An import module is used to construct an initial recommendation model and import user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit;

[0104] An analysis module, configured to perform reliability analysis on the user data through a reliability analysis unit of the initial recommendation model to obtain a reliability value;

[0105] A modeling module, configured to perform user modeling on the user data and the reliability value through a modeling unit of the initial recommendation model to obtain a user feature representation vector, and perform commodity modeling on the user data and the reliability value to obtain a commodity feature representation vector;

[0106] A prediction module, configured to perform a fusion calculation on the user feature representation vector and the product feature representation vector through a score prediction unit of the initial recommendation model to obtain a predicted score;

[0107] An optimization module, configured to calculate the predicted score and the user data through a loss function to obtain a loss value, and optimize the initial recommendation model according to the loss value to obtain a recommendation model;

[0108] The recommendation module is used to calculate the imported to-be-recommended user nodes and to-be-recommended commodity nodes through the recommendation model to obtain commodity recommendation results.

[0109] A product recommendation system provided by an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a product recommendation method as described above is implemented.

[0110] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, a product recommendation method as described above is implemented.

[0111] The above-mentioned product recommendation system and storage medium can refer to the implementation content and beneficial effects of the product recommendation method specifically described above, which will not be repeated here.

[0112] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A product recommendation method, characterized in that: The steps include: Constructing an initial recommendation model, and importing user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit; Performing reliability analysis on the user data by the reliability analysis unit of the initial recommendation model to obtain a reliability value; Performing user modeling on the user data and the reliability value through the modeling unit of the initial recommendation model to obtain a user feature representation vector, and performing product modeling on the user data and the reliability value to obtain a product feature representation vector; The score prediction unit of the initial recommendation model performs a fusion calculation on the user feature representation vector and the product feature representation vector to obtain a predicted score; The predicted score and the user data are calculated by a loss function to obtain a loss value, and the initial recommendation model is optimized according to the loss value to obtain a recommendation model; The recommendation model is used to calculate the imported user nodes to be recommended and the commodity nodes to be recommended to obtain commodity recommendation results.

2. The commodity recommendation method according to claim 1, characterized in that: The user data includes a plurality of user nodes and neighbor sets corresponding to the plurality of user nodes respectively; The reliability analysis unit of the initial recommendation model performs reliability analysis on the user data to obtain a reliability value, specifically: Counting the number of neighbors of the first-order neighbors of the corresponding user node according to the plurality of neighbor sets to obtain a sequence of the number of neighbors of the plurality of user nodes; The plurality of neighbor quantity sequences are respectively converted by encoding using a reliability expression to obtain a context vector corresponding to each of the user nodes, and reliability calculations are performed on the plurality of context vectors to obtain a reliability value corresponding to each of the user nodes. The reliability expression is: Among them, c i is user node u i The context vector of , Encoder is a sequence encoder composed of LSTM units, is user node u i The number of neighbors of the first-order neighbors of user node u i Neighbor Set is user node u i The first-order neighbor of The number of neighbors, r i is user node u i The reliability value of minC is the minimum value of the context vector C of all user nodes, and maxC is the maximum value of the context vector C of all user nodes.

3. The commodity recommendation method according to claim 2, characterized in that: The user data also includes a plurality of commodity nodes and commodity ratings corresponding to the plurality of user nodes; The modeling unit of the initial recommendation model performs user modeling on the user data and the reliability value to obtain a user feature representation vector, specifically: Vectorizing the plurality of user nodes, the neighbor sets corresponding to the plurality of user nodes, the plurality of product nodes, and the product ratings corresponding to the plurality of user nodes, respectively, to obtain a plurality of user embedding vectors, a neighbor embedding vector corresponding to each of the user nodes, a plurality of product embedding vectors, and a rating embedding vector corresponding to each of the user nodes; Performing product aggregation on the plurality of user embedding vectors, the plurality of product embedding vectors, the plurality of reliability values, and the plurality of rating embedding vectors respectively to obtain a user-product embedding vector corresponding to each of the user nodes, wherein the user-product embedding vector corresponding to each of the user nodes includes a neighbor-product embedding vector corresponding to each neighbor node in each neighbor set; Performing social aggregation on the plurality of user embedding vectors and the plurality of neighbor product embedding vectors respectively to obtain a user social embedding vector corresponding to each user node; The user feature expression is used to calculate each user-item embedding vector and the corresponding user-social embedding vector, respectively, to obtain the user feature representation vector corresponding to each user node. The user feature expression is: Among them, h i is user node u i The user feature representation vector is is user node u i The user-item embedding vector of is user node u i The user social embedding vector, W i is the weight matrix.

4. The commodity recommendation method according to claim 3, characterized in that: The performing product aggregation on the plurality of user embedding vectors, the plurality of product embedding vectors, the plurality of reliability values ​​and the plurality of rating embedding vectors to obtain the user product embedding vector corresponding to each user node is specifically as follows: The plurality of reliability values, the plurality of product embedding vectors and the plurality of rating embedding vectors are calculated respectively through an interaction vector expression to obtain an interaction representation vector of each user node for the rated product. The interaction vector expression is: Among them, x ij is user node u i For product node v j The interaction representation vector, f U is a multi-layer perceptron MLP, q j is the commodity node v j The corresponding product embedding vector, e ij is user node u i For product node v j The rating embedding vector corresponding to the rating, r i is user node u i The reliability value of It is the concatenation operation of two vectors; Attention analysis is performed on the multiple user embedding vectors and the corresponding interaction representation vectors through the product embedding expression to obtain the attention corresponding to the multiple interaction representation vectors, and the multiple attentions are allocated to the corresponding interaction representation vectors. The attention allocation results are aggregated to obtain the user product embedding vector corresponding to each user node. The product embedding expression is: Among them, α ij is user node u i For product node v j The attention level, p i is user node u i The corresponding user embedding vector, is user node u i The product embedding vector, x ij is user node u i For product node v j The interaction representation vector is is user node u i The set of evaluated products, ReLU is the activation function, Tanh is the activation function, W I is the weight matrix, b I is the bias vector, T is the matrix transpose, It is the concatenation operation of two vectors.

5. The commodity recommendation method according to claim 3, characterized in that: The social aggregation is performed on the plurality of user embedding vectors and the plurality of neighbor product embedding vectors respectively to obtain the user social embedding vector corresponding to each user node, specifically: Attention analysis is performed on the multiple user embedding vectors and the multiple neighbor product embedding vectors through the social embedding expression to obtain the user social strength corresponding to each user node, and the multiple user social strengths are assigned to the corresponding neighbor product embedding vectors, and the social strength assignment results are aggregated to obtain the user social embedding vector corresponding to each user node. The social embedding expression is: Among them, β ik is user node u i With neighbor node u k The social intensity, is user node u i Corresponding neighbor node u k The neighbor product embedding vector, p i is user node u i The user embedding vector of is user node u i The social embedding vector of is user node u i Neighbor set, ReLU is the activation function, Tanh is the activation function, W s is the weight matrix, b s is the bias vector, is the concatenation operation of two vectors, and T is the matrix transpose.

6. The commodity recommendation method according to claim 3, characterized in that: The product modeling is performed on the user data and the reliability value to obtain a product feature representation vector, specifically: The plurality of reliability values, the plurality of user embedding vectors and the plurality of rating embedding vectors are calculated respectively through the interactive perception vector expression to obtain the interactive perception representation vector of the opinions of each user node on the rated product. The interactive perception vector expression is: Among them, y ji is the commodity node v j With user node u i The opinion interaction perception representation vector, f V is a multi-layer perceptron MLP, p i is user node u i The corresponding user embedding vector, e ij is user node u i For product node v j The rating embedding vector corresponding to the rating, r i is user node u i The reliability value of It is the concatenation operation of two vectors; Attention analysis is performed on the multiple product embedding vectors and the corresponding opinion interaction perception representation vectors through the product feature expression to obtain the attention weights of the multiple opinion interaction perception representation vectors, and the multiple attention weights are assigned to the corresponding opinion interaction perception representation vectors. The attention weight assignment results are aggregated to obtain the product feature representation vector corresponding to each product node. The product feature expression is: Among them, μ ji is the attention weight, q j is the commodity node v j The corresponding product embedding vector, y ji is the commodity node v j With user node u i The opinion interaction perception representation vector, h j is the commodity node v j The feature representation vector of For the node v that has evaluated the product j The user set, ReLU is the activation function, Tanh is the activation function, W j is the weight matrix, b j is the bias vector, T is the matrix transpose, It is the concatenation operation of two vectors.

7. The commodity recommendation method according to claim 6, characterized in that: The scoring prediction unit of the initial recommendation model performs a fusion calculation on the user feature representation vector and the product feature representation vector to obtain a predicted score, specifically: The user feature representation vector corresponding to each user node and the commodity feature representation vector corresponding to each commodity node are calculated respectively through the score calculation expression to obtain the predicted score of each user node for each commodity node. The score calculation expression is: in, is user node u i For product node v j The prediction score, h i is user node u i The user feature representation vector, h j is the commodity node v j The product feature representation vector, W1, W2, W r is the weight matrix, b1 and b2 are bias vectors, Tanh is the activation function, T is the matrix transpose, This is a vector concatenation operation.

8. The commodity recommendation method according to claim 7, characterized in that: The predicted score and the user data are calculated by the loss function to obtain a loss value, and the initial recommendation model is optimized according to the loss value to obtain a recommendation model, specifically: The predicted score and the product score are calculated by using an initial loss function to obtain an initial loss value. The initial loss function is: Among them, L is the initial loss value, r ij is user node u i For product node v j Product ratings, is user node u i For product node v j 's prediction score; The initial recommendation model is iteratively optimized according to the initial loss value through the RMSprop algorithm to obtain a target loss function, and the model parameters of the initial recommendation model are updated according to the target loss function to obtain a recommendation model.

9. A product recommendation system, characterized in that: include: An import module is used to construct an initial recommendation model and import user data collected in advance from the shopping platform into the initial recommendation model, wherein the initial recommendation model includes a reliability analysis unit, a modeling unit, and a rating prediction unit; An analysis module, configured to perform reliability analysis on the user data through a reliability analysis unit of the initial recommendation model to obtain a reliability value; A modeling module, configured to perform user modeling on the user data and the reliability value through a modeling unit of the initial recommendation model to obtain a user feature representation vector, and perform commodity modeling on the user data and the reliability value to obtain a commodity feature representation vector; A prediction module, configured to perform a fusion calculation on the user feature representation vector and the product feature representation vector through a score prediction unit of the initial recommendation model to obtain a predicted score; An optimization module, configured to calculate the predicted score and the user data through a loss function to obtain a loss value, and optimize the initial recommendation model according to the loss value to obtain a recommendation model; The recommendation module is used to calculate the imported to-be-recommended user nodes and to-be-recommended commodity nodes through the recommendation model to obtain commodity recommendation results.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the commodity recommendation method according to any one of claims 1 to 8 is implemented.