Multi-criteria recommendation method and system based on graph representation learning

Abstract

The application relates to the technical field of deep learning recommendation, in particular to a multi-criterion recommendation method and system based on graph representation learning. The method comprises the following steps: constructing a bipartite graph set of an interaction graph; generating a sampled subgraph through a graph sampling strategy, and obtaining a new bipartite graph set for training; initializing graph embedding of nodes in each bipartite graph in the new bipartite graph set; aggregating local embedding and global embedding of each node in the sampled subgraph of each criterion through an aggregation function, and generating embedding representation of each node at each layer under each auxiliary criterion; obtaining embedding representation of each node at each layer under the target criterion according to a multi-head attention mechanism, and further obtaining the final embedding representation of each user and product under the target criterion; updating all parameters to be learned in the model according to an optimization target; and calculating the interaction probability of the user and the product, and generating the order of recommended products for the user. The application improves the training efficiency and precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning recommendation, and particularly relates to a multi-criteria recommendation method and system based on graph representation learning. BACKGROUND

[0002] As one of the important technologies to alleviate the problem of information overload in the Internet era, recommendation systems have been widely applied in e-commerce, social media, online video, and other commercial platforms. The goal of recommendation systems is to mine users' potential behavior patterns from historical interaction data through collaborative filtering or deep learning techniques, and then recommend a small set of products to users. This personalized recommendation technology not only improves the user experience, but also brings economic benefits to online platforms. For example, in the e-commerce field, by analyzing users' purchase history, browsing behavior, search records, and information such as product attributes and descriptions, users are recommended products that they may like, improving the likelihood of user purchases. In the social media field, by analyzing users' social networks, friend relationships, common interests, and other information, as well as users' interaction behaviors on the platform, such as likes, comments, and shares, new friends are recommended. In the online video field, by analyzing users' viewing history, ratings, preferences, and viewing duration, users are recommended films and episodes that they may like, thereby improving the user viewing experience. However, in today's era of rapid development of Internet technology, the rapid increase in the number of users and products poses challenges to recommendation systems in terms of data sparsity and cold start problems. Most traditional recommendation system methods rely on user-product overall rating matrices for modeling, which limits the recommendation system's ability to calculate the similarity of users or products in sparse scenarios. This limitation not only seriously affects the accuracy of personalized recommendation results, but may even harm user experience and activity levels. Therefore, it is crucial to integrate multi-source information to alleviate the challenges posed by data sparsity and cold start problems.

[0003] At present, recommendation methods based on graph representation learning have been widely used in recommendation systems, thereby alleviating the problems of data sparsity and cold start. These methods represent the interaction information between users and products through a graph structure, and use graph neural networks to learn the embedding of nodes in the graph, thereby improving the performance of the recommendation system. However, there are some limitations in existing recommendation methods. On the one hand, the graph representation learning in the recommendation system usually needs to perform message passing and graph convolution operations on the entire graph, which may bring high computational complexity on large-scale graphs. On the other hand, existing methods mostly ignore the importance of multi-criteria rating data for modeling the recommendation system. Even if the modeling is performed using graph representation learning, the recommendation system is still difficult to effectively alleviate the problems of data sparsity and cold start if there is a lack of sufficient supervised information. Multi-criteria information can help the recommendation system to comprehensively capture the inherent characteristics of users and more accurately predict user preferences by extracting complex heterogeneous behavior information. Therefore, it is an urgent need in the field of recommendation systems to develop a computationally efficient recommendation method that can integrate multi-source information and has strong generalization ability. SUMMARY

[0004] The present application overcomes the deficiencies of the prior art. In order to solve the challenges faced by existing recommendation methods in the data sparsity and cold start scenarios, a multi-criteria recommendation method and system based on graph representation learning are proposed. The method improves the training speed of the model on large-scale graph data through a graph sampling strategy, captures the complex relationships between multi-criteria using a graph attention mechanism, improves the generalization ability of the model according to graph contrastive learning, and jointly optimizes the BPR loss and the contrastive loss to improve the robustness of the system in the multi-criteria scenario, thereby alleviating the problems of data sparsity and cold start.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is as follows: a multi-criteria recommendation method based on graph representation learning, comprising the following steps:

[0006] Step S1: obtaining a set of interaction rating data under K criteria and constructing a set of bipartite graphs of the interaction graph according to the interaction rating information of users and products under each criterion The overall rating matrix is defined as the main criterion, and the rest of the rating matrix is defined as the auxiliary criterion.

[0007] Step S2: performing a graph sampling strategy on the bipartite graph corresponding to each criterion to generate its corresponding subgraph, and obtaining a new set of bipartite graphs for training according to the subgraphs corresponding to each criterion

[0008] Step S3: optimizing the new set of bipartite graphs Initialize graph embedding for each node in each bipartite graph to obtain local embedding matrix of user node set and product node set in the 0th layer and global embedding matrix and

[0009] Step S4: aggregate the local embedding and global embedding of each node in each criterion's sampled subgraph by an aggregation function to generate the embedding representation of each node in each layer under K-1 auxiliary criteria, and obtain the embedding representation set of each node respectively represent the embedding representation of node v in the 1th~K-1th layer under the 1th auxiliary criterion obtained by the aggregation function;

[0010] Step S5: obtain the embedding representation of each node in each layer under the target criterion K according to the multi-head attention mechanism Then, according to the embedding representation of each layer of the node under the target criterion K, the final embedding representation of each user and product under the target criterion K is obtained;

[0011] Step S6: update all parameters to be learned in the model according to the optimization objective The optimization objective includes BPR loss and total contrastive loss;

[0012] Step S7: calculate the interaction probability of users and products according to the final embedding representation e u and e i of users and products obtained by the optimized model to generate the order of recommended products for users.

[0013] In the step S2, the specific method of generating the corresponding subgraph by the graph sampling strategy is as follows:

[0014] First, randomly select a plurality of nodes as starting nodes, and generate a random walk sequence {v1, v2,...v T} according to the link information of the neighbors of each starting node, T is the length of the walk, and N groups of walk sequences are generated for each node;

[0015] Then, calculate the frequency of each node appearing in the N groups of walk sequences of each starting node, and find the top M nodes with the highest frequency as the neighbor set of the starting node;

[0016] Finally, generate the corresponding subgraph according to the neighbor set of each starting node.

[0017] The calculation formula of the frequency of each node appearing in the step S2 is as follows:

[0018] p(v t) = sum(v t ) / (N x T).

[0019] p(v t ) represents the probability of the occurrence of node v t , and sum(·) represents the sum of the number of occurrences of node v t .

[0020] In the step S4, the calculation formula of the embedding representation is:

[0021]

[0022] wherein, represents the embedding representation of node v at the l-th layer under the criterion k, respectively represent the embedding representation of node v' at the (l-1)-th layer under the criterion k, represents the set of neighbor nodes of node v, represents the global representation of node v at the (l-1)-th layer, and σ(·) is a LeakyReLU activation function; a v′ represents the attention coefficient of node v and node v';

[0023] or:

[0024]

[0025] wherein, is a weight parameter matrix, is a bias matrix, and Max(·) is used to obtain the maximum value.

[0026] or:

[0027]

[0028] wherein, is a weight parameter matrix, and Mean(·) is used to calculate the average value between vectors.

[0029] In the step S4, the calculation formula of the attention coefficient a v′ is:

[0030]

[0031]

[0032] wherein, W1 is a shared attention mechanism matrix, W2 is a weight parameter matrix, and (·||·) represents a concatenation operation of a matrix and a vector.

[0033] In the step S5, the calculation formula of the embedding representation of each node at each layer under the criterion K is:

[0034]

[0035] in, and Let A and B represent the embedding representations of node v in the l-th and (l-1)-th layers under the target criterion K, respectively. Let A represent the number of attention mechanism transformations, and β represent the hyperparameter used to control the intensity of auxiliary criterion features injected into the target criterion. Attention coefficient This represents the set of embedding representations of node v at the (l-1)th layer under the 1st to (k-1)th auxiliary criteria;

[0036] The formula for calculating the attention coefficient is:

[0037]

[0038] Where Softmax(·) represents the activation function used for normalization; tanh(·) represents the hyperbolic tangent function; and Let be the parameter matrix to be learned; Let v represent the set of embedding representations of node v at the l-th layer under the 1st to K-1th auxiliary criteria.

[0039] In step S5, the calculation formula for the final embedded representation of each user and product under the target criterion K is as follows:

[0040]

[0041] Among them, e u With e i Let represent the final embedding representations of user u and product i under the target criterion K, respectively. and Let u and i represent the embedding representations of user u and product i at the l-th layer under the target criterion K, respectively.

[0042] In step S6, the optimization target The calculation formula is:

[0043]

[0044] in, Indicates BPR loss, Let λ1 represent the total contrastive loss, λ2 be the hyperparameter, and λ2 be the regularization parameter. The parameter Θ to be learned is regularized using the L2 norm;

[0045] BPR loss The calculation formula is:

[0046]

[0047] wherein, and respectively represent the BPR loss under the objective criterion K and the kth auxiliary criterion, and the calculation formula is:

[0048]

[0049] wherein, (u,j) represents a set of negative example pairs randomly sampled;

[0050] respectively represent the predicted value of the interaction probability between the user u and the products i and j, represent the positive and negative sample sets; represent the sampled subgraph under the kth sampling criterion;

[0051] the total contrast loss The calculation formula is:

[0052]

[0053] wherein, represent the contrast loss between the objective criterion K and the 1st, 2nd, … kth, …, K-1th auxiliary criterion; The calculation formula is:

[0054]

[0055] wherein, is the negative example set, and s(·) is used to calculate the similarity between vectors; τ is a hyperparameter used to control the contrast intensity; respectively represent the embedding representation of node v under the objective criterion K and the kth auxiliary criterion, represent the embedding representation of node v' under the kth auxiliary criterion, represent the node set of the user and the product.

[0056] In the step S7, the calculation formula of the interaction probability between the user and the product is:

[0057]

[0058] represent the predicted value of the interaction probability between the user u and the product i.

[0059] In addition, the present application also provides a multi-criteria recommendation system based on graph representation learning, comprising:

[0060] A bipartite graph construction module is used to obtain the interaction score data set under K criteria and construct a bipartite graph set of interaction graphs according to the interaction score information of the user and the product under each criterion The total score matrix is defined as the objective criterion, and the rest of the score matrix is defined as the auxiliary criterion. The total score matrix is defined as the objective criterion, and the rest of the score matrix is defined as the auxiliary criterion. The total score matrix is defined as the objective criterion, and the rest of the score matrix is defined as the auxiliary criterion.

[0061] The graph sampling module is used to perform a graph sampling strategy on the bipartite graph corresponding to each criterion to generate a corresponding subgraph, and a new bipartite graph set for training is obtained according to the subgraph corresponding to each criterion.

[0062] The graph embedding initialization module is used to perform graph embedding initialization on the nodes in each bipartite graph in the new bipartite graph set, and obtain the local embedding matrix of the user node set and the product node set in the 0th layer. And And And the global embedding matrix And

[0063] The single-criterion information propagation module is used to aggregate the local embedding and the global embedding of each node in the subgraph of each criterion through an aggregation function, to generate the embedding representation of each node in each layer under K-1 auxiliary criteria, and obtain the embedding representation set of each node.

[0064] The multi-criterion feature aggregation module is used to obtain the embedding representation of each node in each layer under the objective criterion K according to a multi-head attention mechanism. Then, according to the embedding representation of each layer of the node under the objective criterion K, the final embedding representation of each user and product under the objective criterion K is obtained.

[0065] The joint optimization module is used to update all parameters to be learned in the model according to an optimization objective. The optimization objective includes a BPR loss and a total contrastive loss.

[0066] The prediction output module is used to calculate the interaction probability of the user and the product according to the final embedding representation e u of the user and the product obtained by the optimized model, and generate a sequence of recommended products for the user. i

[0067] Compared with the prior art, the present application has the following beneficial effects:

[0068] ​1. The application adopts a graph sampling strategy in a multi-criteria recommendation scenario, reduces multiple interactions and complex topological structures in graph data, and the purpose of sampling is to convert large graph data into controllable small batch size by randomly extracting subgraphs, thereby improving the training efficiency of the model and reducing the complexity of calculation under the multi-criteria scene, saving the huge calculation and memory overhead generated by directly processing such large-scale graph data, and making the training process simple and efficient;

[0069] 2. The application constructs a graph representation learning architecture for a multi-criteria recommendation scenario, utilizes high-order connectivity to recursively propagate local embedding and global embedding, and the architecture is divided into single-criteria feature aggregation and multi-criteria information propagation two parts to extract collaborative filtering signals between nodes and criteria; the core of single-criteria feature aggregation is to aggregate feature information from the sampling field according to the graph convolution operation to update the embedding representation under the specific criteria interaction graph; multi-criteria information propagation enhances the embedding performance under the target criteria score through cross-view dependency modeling;

[0070] 3. The application constructs self-supervised contrastive learning and multi-criteria view interaction prediction to improve the robustness of the model, so that the recommendation system can extract rich user behavior preferences to alleviate the data sparsity and cold start problem;

[0071] 4. By jointly optimizing the contrastive loss and the prediction loss, in four real-world multi-criteria recommendation scenarios, such as the movie recommendation dataset Yahoo! Movie, the hotel recommendation dataset TripAdvisor, the beer recommendation dataset RateBeer, and the commercial place recommendation dataset Yelp-2022, compared with the current best method CPA-LGC, the application respectively obtains the best performance in the precision and recall two indicators; on the Yahoo! Movie dataset, the two indicators are improved by 4.32% and 3.84% respectively; on the TripAdvisor dataset, the two indicators are improved by 3.61% and 6.39% respectively; on the RateBeer dataset, the two indicators are improved by 4.35% and 2.67% respectively; on the Yelp-2022 dataset, the two indicators are improved by 5.18% and 4.54% respectively. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 A flowchart of a multi-criteria recommendation method based on graph representation learning provided for example one;

[0073] Figure 2 A schematic diagram of the principle of the graph sampling strategy;

[0074] Figure 3 A single-criteria information propagation diagram;

[0075] Figure 4 a multi-criteria recommendation system based on graph representation learning;

[0076] Figure 5 a structure diagram of a multi-criteria recommendation system based on graph representation learning provided for embodiment two. DETAILED DESCRIPTION

[0077] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0078] As shown in Figures 1-3 , the embodiments of the present application provide a multi-criteria recommendation method based on graph representation learning, comprising the following steps:

[0079] Step S1: obtaining an interaction score data set under K criteria and constructing a bipartite graph set of the interaction graph according to the interaction score information of the user and the product under each criterion The total score matrix is defined as the main criterion, and the rest of the score matrix is defined as the auxiliary criterion.

[0080] Generally, in the recommendation system scenario, assuming that the user has K criteria scores for the product, a multi-criteria recommendation scenario can be defined;

[0081] wherein the obtained multi-criteria score data set is:

[0082]

[0083] wherein, the matrix element in the score matrix corresponding to the kth criterion is , which represents the score value of the user u for the product i under the criterion k is c, , which represents the size of the matrix is and respectively represent the user set and the product set, and are the number of users and products, respectively. Further, the total score matrix is set as the main criterion corresponding to the criterion, and the K-1 score matrices are set as auxiliary criteria corresponding to the criteria.

[0084] Wherein, the bipartite graph set for constructing the interaction graph is:

[0085]

[0086] Wherein, wherein is a node set, and ε is an edge set. If the real rating value of the user u to the product i in the criterion k is greater than the median value, there is an edge connecting the user u and the product i in the graph formed by the criterion k.

[0087] Step S2: performing a graph sampling strategy on each bipartite graph corresponding to each criterion to generate a corresponding subgraph, and obtaining a new bipartite graph set for training according to the subgraph corresponding to each criterion

[0088] The graph sampling strategy can adaptively process graph data of different structures and scales, and can ensure the universality and scalability of the model in various multi-criteria recommendation scenarios. The neighborhood set generated by the graph sampling strategy is not limited to the first-order neighbor, and the graph construction method can select more important high-order neighbors for the node, which can propagate key features in a more refined manner. The graph sampling strategy is an inductive learning paradigm, which has the ability to generalize to unknown nodes, so that the model can effectively deal with the cold start problem. In the embodiment, the graph sampling strategy is performed on each bipartite graph corresponding to each criterion, and each graph is generated to generate a corresponding subgraph, so that the recommendation system adaptively processes graph data of different structures and scales, and converts the huge graph data into a controllable small batch size, improving the operation speed.

[0089] In the step S2, the specific method for performing the graph sampling strategy to generate the corresponding subgraph is:

[0090] (1) First, randomly select a plurality of nodes as starting nodes, and generate a random walk sequence according to the link information of the neighbors of each starting node. Taking the kth criterion as an example, select one node v, and generate a random walk sequence according to the link information of the neighbors of the node v, which is {v1, v2,...v T}, T is the length of the walk, and N groups of walk sequences are generated for the node v.

[0091] (2) Then, calculate the frequency of each node appearing in the N groups of walk sequences of each starting node, find the top M nodes with the highest frequency, and take them as the neighbor set of the starting node Wherein, The calculation formula of the frequency of each node is:

[0092] p(v t )=sum(v t ) / (N×T); (3)

[0093] p(v​t ) represents the node v t the probability of occurrence, sum(·) represents the node v t the sum of the number of occurrences.

[0094] (3) Finally, according to the neighbor set of each starting node, the corresponding subgraph is generated for subsequent training.

[0095] As Figure 2 shown in the principle diagram of the graph sampling strategy. The bipartite graph set generated by the graph sampling strategy for training is:

[0096]

[0097] Among them, is the subgraph under criterion k.

[0098] Step S3: Perform graph embedding initialization on the nodes in each bipartite graph in the new bipartite graph set , to obtain the local embedding matrices of the user node set and the product node set in the 0th layer and and the global embedding matrices and

[0099] Graph embedding technology is a method of mapping complex graph data to a low-dimensional vector space, which has been widely applied in the field of recommendation systems. By mining high-order topological structure features, the recommendation algorithm can more accurately capture personalized user interests, thereby improving the accuracy of model recommendations. Different criteria may have certain differences, and local embedding is used to capture user-specific behavior preferences under each criterion, so that the system can better understand the data characteristics under each criterion. Different criteria may have certain correlations, and step S3 introduces a set of global embeddings as shared vector representations between all criteria. The purpose of this strategy is to build a bridge for cross-criterion information transmission, thereby enhancing the model's ability to integrate and learn global information and improving the system's generalization performance in a multi-criteria scenario.

[0100] As Figure 3 shown in the single-criterion information propagation diagram. Among them, the specific method of step S3 is:

[0101] (1) The local graph embedding initialization method is:

[0102]

[0103] Among them, and represent the feature matrices of users and products in the 0th layer in the rating under criterion k, and the features are where d is the size of the embedding dimension, and T is the transpose operation of the matrix; u m and i m respectively represent the mth user u and product i, both of which are one-hot vectors to ensure the uniqueness of each node after initialization; and respectively represent the mth user u and product i in the 0th layer under criterion k; feature matrix and After initialization, the embedding features of each user and product in the 0th layer under each criterion can be extracted according to the above formula.

[0104] (2) Global embedding matrix and The set of

[0105]

[0106] where and respectively represent the initialized set of global embeddings of users and products, and respectively represent the global embedding representation of the mth user and product in the 0th layer.

[0107] Step S4: Aggregate the local embedding and global embedding of each node in the sampled subgraph of each criterion by using an aggregation function to generate the embedding representation of each node in each layer under K-1 auxiliary criteria, and obtain the set of embedding representations of each node respectively represent the embedding representation of node v in the 1st to K-1 auxiliary criteria in the lth layer obtained by the aggregation function.

[0108] The core of single-criterion feature aggregation is the graph convolution operation, that is, how to aggregate feature information from the sampling field of node v to update the embedding representation under the specific criterion interaction graph; in actual scenarios, a targeted aggregation method can be designed according to the specific actual scenario.

[0109] The embodiment of the application adopts multi-criterion feature aggregation, as shown in Figure 4 The principle diagram of multi-criterion feature aggregation is shown.

[0110] In this embodiment, the calculation formula of the embedding representation of the node is:

[0111]

[0112] where is the sampled subgraph under criterion k The neighborhood set of the middle node v, σ(·) is a LeakyReLU activation function, f aggregator (·) represents an aggregation function, represents the embedding representation of the neighbor v' of the node v at the l-1 layer, represents the global embedding of the node v at the l-1 layer.

[0113] The embodiment of the application can update the embedding representation of the node through the following three aggregation functions.

[0114] First, the first is an attention aggregation function, and the calculation formula is:

[0115]

[0116] wherein, represents the embedding representation of the node v at the l layer under the criterion k, respectively represent the embedding representation of the node v' at the l-1 layer under the criterion k, represents the set of neighbor nodes of the node v, represents the global representation of the node v at the l-1 layer, σ(·) is a LeakyReLU activation function; a v′ represents the attention coefficient of the node v and the node v'; the attention coefficient a v′ The calculation formula is:

[0117]

[0118]

[0119] wherein, W1 is a shared attention mechanism matrix, W2 is a weight parameter matrix, and (·||·) represents a splicing operation of a matrix vector. d' represents the dimension of the matrix.

[0120] The second is a max-pooling aggregation function, which accepts the node features with the largest influence in the local domain, thereby retaining the most significant embedding representation in the neighborhood, and the specific calculation method is:

[0121]

[0122] wherein, is a weight parameter matrix, is a bias matrix, and Max(·) is used to obtain the maximum value;

[0123] The third is an average pooling, which extracts the average value in the local domain as the output, thereby obtaining the node features of the overall neighborhood to update the embedding representation, and the calculation method is:

[0124]

[0125] wherein, is a weight parameter matrix, Mean(·) is used to calculate the average value between vectors;

[0126] Finally, for any node v, the embedding representation set of v under the K-1 criteria is represented as wherein, respectively represent the embedding representation of the l-th layer of v under the k-th criterion obtained by the aggregation function.

[0127] Step S5: obtaining the embedding representation of each node under the K-th criterion at each layer according to the multi-head attention mechanism Then, according to the embedding representation of each layer of the node under the K-th criterion, the final embedding representation of each user and product under the K-th criterion is obtained.

[0128] In this embodiment, the correlation between the K-th criterion and the auxiliary criteria is measured according to the multi-head attention mechanism, the interaction behavior between different criteria is modeled to enhance the embedding features in the K-th criterion, and the final vector representation of the node is obtained after aggregation. The multi-criteria information propagation enhances the embedding performance under the K-th criterion score by modeling the dependence across views.

[0129] In view of the preference difference of users between different criteria, in the step S5, the graph attention mechanism is introduced to measure the correlation between the K-th criterion and the auxiliary criteria, and the following multi-head attention is performed to aggregate the embedding on different criteria to enhance the node representation in the K-th criterion.

[0130] In the step S5, the embedding representation of the user and the product is obtained according to the multi-head attention mechanism, and the calculation formula of the embedding representation of each node under the K-th criterion at each layer is as follows:

[0131]

[0132] wherein, and respectively represent the embedding representation of the l-th and l-1-th layer of the node v under the K-th criterion, A represents the number of attention mechanism conversions, β represents a hyperparameter for controlling the strength of the auxiliary criterion features injected into the K-th criterion; is an attention coefficient, which reflects the dependence relationship of the node v (user or product) in the K-th criterion with respect to its K-1 criteria, The greater the value of, the greater the influence of the embedding features of the K-th criterion on the node v, denotes the embedding representation set of node v at the l-1th layer under the 1th to K-1th auxiliary criteria. wherein the calculation formula of the attention coefficient is:

[0133]

[0134] wherein Softmax(·) denotes an activation function for normalization processing; tanh(·) denotes a hyperbolic tangent function, which maps the input value to the interval (-1, 1), so that the network structure is more stable and reliable; and is a parameter matrix to be learned. denotes the embedding representation set of node v at the lth layer under the 1th to K-1th auxiliary criteria.

[0135] On this basis, the node features on each layer criterion are aggregated, and the calculation is as follows:

[0136]

[0137] wherein e u and e i respectively denote the final embedding representation of user u and product i on the criterion K, and respectively denote the embedding representation of user u and product i at the lth layer under the criterion K.

[0138] Step S6: updating all parameters to be learned in the model according to the optimization target ; the optimization target includes the BPR loss and the total contrastive loss.

[0139] The contrastive loss maps similar users or products to adjacent spatial regions, which helps to improve the understanding of the complex relationship between users and products by the system, alleviates the data sparsity and cold start problem, and thus better performs personalized recommendation; the BPR loss is a loss function used for ranking tasks in a recommendation system, which better understands the user's preferences by maximizing the ranking difference between positive examples and negative examples. Specifically, in the step S6, the calculation formula of the optimization target is:

[0140]

[0141] wherein, denotes the BPR loss, denotes the total contrastive loss, λ1 is a hyperparameter, and λ2 is a regularization parameter, denotes the regularization of the parameter Θ to be learned by the L2 norm.

[0142] BPR loss The calculation formula is:

[0143]

[0144] Wherein, γ is a hyperparameter, used to control the influence strength of the auxiliary criterion loss on the model; And Respectively represent the BPR loss under the target criterion K and the kth auxiliary criterion, and the calculation formula is:

[0145]

[0146] Wherein, (u,j) represents a set of negative example pairs randomly sampled, used to improve the calculation efficiency during model training; Respectively represent the predicted value of the interaction probability between user u and products i and j, Respectively represent the positive and negative sample sets; Respectively represent the sampled subgraph under the kth sampling criterion;

[0147] Total contrastive loss The calculation formula is:

[0148]

[0149] Wherein, Respectively represent the contrastive loss between the target criterion K and the 1st, 2nd, … kth, …, K-1th auxiliary criterion; The calculation formula is:

[0150]

[0151] Wherein, Is the negative example set, s(·) is used to calculate the similarity between vectors; τ is a temperature coefficient, a hyperparameter used to control the contrast strength; Respectively represent the embedding representation of node v under the target criterion K and the kth auxiliary criterion, Respectively represent the embedding representation of node v' under the kth auxiliary criterion, Respectively represent the node set of users and products.

[0152] Step S7: According to the final embedding representation e u And e i Of users and products obtained by the optimized model, calculate the interaction probability of users and products, and generate the order of recommended products for users.

[0153] In the step S7, the calculation formula of the interaction probability of users and products is:

[0154]

[0155] is a predicted value representing the interaction probability between user u and product i, which represents the potential preference of u to i. In this embodiment, the final representation e u and e i feeds the prediction function of the above formula to predict the interaction probability between u and i.

[0156] Embodiment Two

[0157] As shown in the figure, the embodiment two of the present application provides a multi-criteria recommendation system based on graph representation learning, comprising: Figure 5 A bipartite graph construction module is configured to obtain a set of interaction score data sets under K criteria

[0158] and construct a set of bipartite graphs of interaction graphs according to the interaction score information of users and products under each criterion The overall score matrix is defined as the main criterion, and the rest of the score matrix is defined as the auxiliary criterion.

[0159] A graph sampling module is configured to generate a corresponding subgraph by performing a graph sampling strategy on the bipartite graph corresponding to each criterion, and obtain a new set of bipartite graphs for training according to the subgraph corresponding to each criterion

[0160] A graph embedding initialization module is configured to perform graph embedding initialization on the nodes in each bipartite graph in the new set of bipartite graphs to obtain the local embedding matrix of the user node set and the product node set in the 0th layer and and the global embedding matrix and

[0161] A single-criterion information propagation module is configured to aggregate the local embedding and the global embedding of each node in the subgraph of each criterion by an aggregation function to generate the embedding representation of each node at each layer under K-1 auxiliary criteria, and obtain a set of embedding representations of each node.

[0162] A multi-criteria feature aggregation module is configured to obtain the embedding representation of each node at each layer under the main criterion K according to a multi-head attention mechanism and then obtain the final embedding representation of each user and product under the main criterion K according to the embedding representation of each layer of the node under the main criterion K.

[0163] A joint optimization module is configured to update all parameters to be learned in the model according to an optimization objective ; the optimization objective ​The BPR loss and the total contrastive loss are included.

[0164] The prediction output module is configured to obtain the final embedding representation e of the user and the product according to the optimized model. u The interaction probability of the user and the product is calculated, and a sequence of recommended products is generated for the user. i The interaction probability of the user and the product is calculated, and a sequence of recommended products is generated for the user.

[0165] In addition, the application also provides a computer device / system, which comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to the first embodiment.

[0166] In addition, the application also provides a computer readable storage medium, which stores a computer program / instruction, wherein the computer program / instruction is executed by a processor to implement the steps of the method according to the first embodiment.

[0167] Further, the application also provides a computer program product, which comprises a computer program, wherein the computer program is executed by a processor to implement the steps of the method according to the first embodiment.

[0168] Those skilled in the art should also know that, in addition to implementing the client and the server in the form of pure computer readable program code, the same functions can also be implemented by logically programming the method steps in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers. Therefore, such a client and server can be considered as a hardware component, and the devices included therein for implementing various functions can also be considered as structures in the hardware component. Alternatively, the devices for implementing various functions can be considered as both software modules for implementing the method and structures in the hardware component.

[0169] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and the necessary general hardware platform. Based on such an understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0170] The various embodiments in the specification are described in progressive manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, the embodiments of the computer storage medium, the server and the client can be explained by referring to the description of the embodiments of the method.

[0171] The present application can be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The present application can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules can be located in local and remote computer storage media including memory storage devices.

[0172] Finally, it should be noted that the above-described embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the present application; even though the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-criteria recommendation method based on graph representation learning, characterized in that, Includes the following steps: Step S1: Obtain the set of interactive rating data under K criteria. And based on the user and product interaction rating information under each criterion, a bipartite graph set of interaction diagrams is constructed. Total score matrix Defined as the target criterion, the remaining scoring matrix Defined as an auxiliary criterion; Step S2: Generate a corresponding sampling subgraph on the bipartite graph corresponding to each criterion using a graph sampling strategy. Obtain a new set of bipartite graphs for training based on the sampling subgraphs corresponding to each criterion. Step S3: For the new bipartite graph set Graph embedding initialization is performed on each node in the bipartite graph to obtain the local embedding matrix of the user node set and the product node set in layer 0. and and global embedding matrix and Step S4: Aggregate the local and global embeddings of each node in the sampling subgraph of each criterion using an aggregation function to generate the embedding representation of each node at each layer under K-1 auxiliary criteria, thus obtaining the embedding representation set of each node. Let each represent the embedding representation of node v at the l-th layer under the 1st to K-1 auxiliary criteria obtained by the aggregation function; Step S5: Obtain the embedding representation of each node at each layer under the target criterion K based on the multi-head attention mechanism. Then, based on the embedding representation of each node under the target criterion K, the final embedding representation of each user and product under the target criterion K is obtained. Step S6: Based on the optimization objective To update all the parameters to be learned in the model; the optimization objective Including BPR loss and total contrast loss; Step S7: Obtain the final embedded representation of users and products based on the optimized model. u With e i Calculate the probability of user interaction with products and generate a recommended product order for the user.

2. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S2, the specific method for generating the corresponding subgraph using the graph sampling strategy is as follows: First, randomly select several nodes as starting nodes, and generate a random walk sequence {v1, v2, ... v3} based on the link information of the neighbors of each starting node. T }, where T is the length of the walk, and N walk sequences are generated for each node; Next, calculate the frequency of each node in the N sets of walk sequences for each starting node, and find the top M nodes with the highest frequency to form the neighbor set of the starting node. Finally, based on the neighbor set of each starting node, the corresponding subgraph is generated.

3. The multi-criteria recommendation method based on graph representation learning according to claim 2, characterized in that, In step S2, the formula for calculating the frequency of each node is as follows: p(v t )=sum(v t ) / (N×T); p(v t ) represents node v t The probability of occurrence, sum(·) represents the probability of node v. t The total number of times it appears.

4. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S4, the calculation formula for the embedding representation is: in, This represents the embedding representation of node v at the l-th layer under criterion k. Let represent the embedding representation of node v' at the (l-1)th layer under criterion k, respectively. Let v represent the set of neighboring nodes of node v. This represents the global representation of node v at layer l-1, where σ(·) is the Leaky ReLU activation function; a v′ This represents the attention coefficient between node v and node v'; or: in, The weight parameter matrix, For the bias matrix, Max(·) is used to obtain the maximum value; or: in, Given the weight parameter matrix, Mean(·) is used to calculate the average value between the vectors.

5. The multi-criteria recommendation method based on graph representation learning according to claim 4, characterized in that, In step S4, the attention coefficient a v′ The calculation formula is: Where W1 is the shared attention mechanism matrix, W2 is the weight parameter matrix, and (·||·) represents the matrix-vector concatenation operation.

6. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S5, the calculation formula for the embedding representation of each node at each layer under the target criterion K is as follows: in, and Let A and B represent the embedding representations of node v in the l-th and (l-1)-th layers under the target criterion K, respectively. Let A represent the number of attention mechanism transformations, and β represent the hyperparameter used to control the intensity of auxiliary criterion features injected into the target criterion. Attention coefficient This represents the set of embedding representations of node v at the (l-1)th layer under the 1st to (k-1)th auxiliary criteria; The formula for calculating the attention coefficient is: Where Softmax(·) represents the activation function used for normalization; tanh(·) represents the hyperbolic tangent function; and Let be the parameter matrix to be learned; Let v represent the set of embedding representations of node v at the l-th layer under the 1st to K-1th auxiliary criteria.

7. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S5, the calculation formula for the final embedded representation of each user and product under the target criterion K is as follows: Among them, e u With e i Let represent the final embedding representations of user u and product i under the target criterion K, respectively. and Let u and i represent the embedding representations of user u and product i at the l-th layer under the target criterion K, respectively.

8. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S6, the optimization target The calculation formula is: in, Indicates BPR loss, λ represents the total contrastive loss, λ1 is a hyperparameter, and λ2 is a regularization parameter. The parameter Θ to be learned is regularized using the L2 norm; BPR loss The calculation formula is: in, and Let K and the k-th auxiliary criterion represent the BPR loss, respectively, and their calculation formulas are as follows: in, (u,j) represents a set of negative pairs randomly sampled; Let represent the predicted probabilities of interaction between user u and products i and j, respectively. Represents the set of positive and negative samples; This represents the sampling subgraph under the k-th sampling criterion; Total comparative loss The calculation formula is: in, This represents the contrast loss between the target criterion K and the 1st, 2nd, ..., kth, K-1th auxiliary criteria; The calculation formula is: in, Let be the set of negative examples, s(·) is used to calculate the similarity between vectors; τ is a hyperparameter used to control the contrast intensity; Let K and K represent the embedding representations of node v under the target criterion K and the k-th auxiliary criterion, respectively. This represents the embedding representation of node v' under the k-th auxiliary criterion. This represents the set of nodes representing users and products.

9. The multi-criteria recommendation method based on graph representation learning according to claim 1, characterized in that, In step S7, the formula for calculating the probability of user-product interaction is: This represents the predicted probability of interaction between user u and product i.

10. A multi-criteria recommendation system based on graph representation learning, characterized in that, include: Bipartite graph construction module: used to obtain a set of interactive rating data under K criteria. And based on the user and product interaction rating information under each criterion, a bipartite graph set of interaction diagrams is constructed. Total score matrix Defined as the target criterion, the remaining scoring matrix Defined as an auxiliary criterion; Graph sampling module: Used to generate a corresponding sampling subgraph on the bipartite graph corresponding to each criterion, and obtain a new set of bipartite graphs for training based on the sampling subgraphs corresponding to each criterion. Graph embedding initialization module: used for initializing new bipartite graph sets. Graph embedding initialization is performed on each node in the bipartite graph to obtain the local embedding matrix of the user node set and the product node set in layer 0. and and global embedding matrix and Single-criterion information propagation module: used to aggregate the local and global embeddings of each node in the sampling subgraph of each criterion through the aggregation function, generate the embedding representation of each node at each layer under K-1 auxiliary criteria, and obtain the embedding representation set of each node; Multi-criteria feature aggregation module: used to obtain the embedding representation of each node at each layer under the target criterion K according to the multi-head attention mechanism. Then, based on the embedding representation of each node under the target criterion K, the final embedding representation of each user and product under the target criterion K is obtained. Joint optimization module: used to optimize based on the optimization objective To update all the parameters to be learned in the model; the optimization objective Including BPR loss and total contrast loss; Prediction Output Module: Used to generate the final embedded representations of users and products based on the optimized model. u With e i Calculate the probability of user interaction with products and generate a recommended product order for the user.

Images