A Graph Neural Network Recommendation Method Integrating Multiple User Behaviors

Abstract

The present invention discloses a graph neural network recommendation method that integrates multiple user behaviors, and uses supervised learning and self-supervised learning to improve the recommendation performance of the graph neural network that integrates multiple user behaviors; in supervised learning, the method of graph convolution is used to learn the historical interaction information of multiple user behaviors, capture the high-order connectivity and structural connectivity between users and items, and obtain high-quality embedding representations of users and items; in self-supervised learning, by performing contrastive learning between the target behavior and other auxiliary behaviors and contrastive learning tasks for individual behaviors, more information can be learned for representation learning, improving the generalization ability and recommendation accuracy of the graph neural network that integrates multiple user behaviors; in addition, by controlling the normalization coefficient in the graph convolution process, it is possible to ensure that the recommendation system improves the novelty of recommendations without losing recommendation accuracy, and alleviates the popularity bias.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and specifically relates to a graph neural network recommendation method that fuses multiple user behaviors. Background Art

[0002] With the development of information technology and the Internet industry, people's demand for and dependence on the Internet are also growing rapidly. The rapid development of the Internet has brought convenience to people's lives, but at the same time has led to the problem of information overload. Facing the vast amount of data on the Internet, it is difficult for users to find the content they need from the vast amount of information within an effective time, and it has become more difficult for merchants to quickly attract users' attention and accurately grasp users' needs with limited resources. As an important method to solve information overload, the recommendation system provides personalized recommendation information that meets users' needs by analyzing users' historical behaviors, preferences, and other information. This can help users find interesting content more quickly and improve the user experience. However, traditional recommendation systems are affected by the following two problems in their recommendation performance: on the one hand, the datasets of traditional recommendation systems all have the problems of data sparsity and long-tail distribution; on the other hand, traditional recommendation systems are designed for a single type of behavior and mainly focus on the interaction between users and items, paying less attention to the specific behaviors in the interaction, while in real scenarios, users usually have more than one type of behavior. Summary of the Invention

[0003] The problem to be solved by the present invention is the problems existing in traditional recommendation systems, and provides a graph neural network recommendation method that fuses multiple user behaviors.

[0004] To solve the above problems, the present invention is implemented through the following technical solutions:

[0005] A graph neural network recommendation method that fuses multiple user behaviors includes the following steps:

[0006] Step 1: Download publicly available multi-behavior data from the Internet, preprocess the multi-behavior data to obtain a multi-behavior dataset, and divide the multi-behavior data into a multi-behavior training set and a multi-behavior test set;

[0007] Step 2: Use the multi-behavior training set to train the graph neural network that fuses multiple user behaviors, and optimize the comprehensive loss of the trained graph neural network that fuses multiple user behaviors by adjusting all trainable parameters of the trained graph neural network that fuses multiple user behaviors, to obtain the trained graph neural network that fuses multiple user behaviors, that is:

[0008] Step 2.1: Use the graph neural network that fuses multiple user behaviors to obtain the embedding representations of users and items under each behavior, and merge the embedding representations of users and items under all behaviors to obtain the final representations of users and items;

[0009] Step 2.2: Conduct supervised learning on the graph neural network that fuses multiple user behaviors, and calculate the loss of the supervised learning task using ranking loss;

[0010] Step 2.3: Conduct the first self-supervised contrast learning on the graph neural network that fuses multiple user behaviors, that is, conduct contrast learning between the target behavior and each auxiliary behavior, and calculate the loss of the first self-supervised contrast learning task using contrast loss;

[0011] Step 2.4: Conduct the second self-supervised contrast learning on the graph neural network that fuses multiple user behaviors, that is, conduct contrast learning for a single behavior, and calculate the loss of the second self-supervised contrast learning task using contrast loss;

[0012] Step 2.5: Combine the losses in Steps 2.2 - 2.4, and optimize the comprehensive loss of the graph neural network that fuses multiple user behaviors by adjusting the trainable parameters of the graph neural network that fuses multiple user behaviors to obtain a trained graph neural network that fuses multiple user behaviors; where the comprehensive loss is:

[0013]

[0014] Among them, is the loss of the supervised learning task, is the loss of the first self-supervised contrast learning task, is the loss of the second self-supervised contrast learning task, λ1 is the weight of the loss of the first self-supervised contrast learning task, λ2 is the weight of the loss of the second self-supervised contrast learning task, Θ is all the trainable parameters of the graph neural network that fuses multiple user behaviors, μ is the hyperparameter of L2 regularization, represents L2 regularization;

[0015] Step 3: Use the multi-behavior test set to test the trained graph neural network that fuses multiple user behaviors, and optimize the evaluation metrics of the accuracy and novelty of the trained graph neural network that fuses multiple user behaviors by adjusting the graph convolution normalization coefficient of the trained graph neural network that fuses multiple user behaviors to obtain the final graph neural network that fuses multiple user behaviors;

[0016] Step 4: Use the final graph neural network that fuses multiple user behaviors to predict the recommended items for the current user, and recommend the prediction results to the current user.

[0017] In the above Step 2.1, the embedding representations of the user under all behaviors are merged using the fusion operation to obtain the final representation of the user; the final representation e u of user u is:

[0018]

[0019] Among them, e uk is the embedded representation of user u under behavior k, a uk is the semantic fusion coefficient of the k-th behavior of user u, K is the total number of behaviors, σ(■) represents the Relu activation function, and W and b are the weights and biases in the fully connected layer of the graph neural network that fuses the multiple behaviors of the user.

[0020] The semantic fusion coefficient a of the k-th behavior of user u uk is:

[0021]

[0022] Among them, w k is the intensity weight of behavior k, n uk is the number of interactions that user u has under behavior k, and K is the total number of behaviors.

[0023] In step 2.1 above, the embedded representations of items under all behaviors are merged using the concatenation operation to obtain the final representation of the item; the final representation e of item i i is:

[0024] e i = g{Cat(e ik )}

[0025] Among them, e ik is the embedded representation of item i under each behavior k, Cat(■) represents the concatenation operation of the graph neural network that fuses the multiple behaviors of the user, and g(■) represents the multi-layer perception operation of the graph neural network that fuses the multiple behaviors of the user.

[0026] In step 2.2 above, the loss of the supervised learning task is:

[0027]

[0028] Among them, O = {(u, i, j)∣(u, i) ∈ O + , (u, j) ∈ O -} is the multi-behavior training set, O + represents the set of users and items that have interactions in the multi-behavior training set, O - represents the set of users and items that have no interactions in the multi-behavior training set, e i is the final representation of item i, e j is the final representation of item j, e u is the final representation of user u, T represents the transpose, and σ(■) represents the Relu activation function.

[0029] In step 2.3, when performing the first self-supervised contrastive learning: First, based on the characteristics of the multi-behavior training set, one of all the behaviors is designated as the target behavior, and the remaining behaviors are designated as auxiliary behaviors; then, the user contrast view pairs are constructed using the embedded representations of users under the target behavior and the embedded representations of users under other auxiliary behaviors to achieve the self-supervised contrastive learning of users. At the same time, the item contrast view pairs are constructed using the embedded representations of items under the target behavior and the embedded representations of items under other auxiliary behaviors to achieve the self-supervised contrastive learning of items.

[0030] In step 2.3, the loss of the first self-supervised contrastive learning task is:

[0031]

[0032] Where is the positive user pair of the user contrast view pair in the first self-supervised contrastive learning task, is the negative user pair of the user contrast view pair in the first self-supervised contrastive learning task, is the behavior the embedded representation of user u under, e uk is the embedded representation of user u under behavior k, e vk is the embedded representation of user v under behavior k; is the positive item pair of the item contrast view pair in the first self-supervised contrastive learning task, is the negative item pair of the item contrast view pair in the first self-supervised contrastive learning task, is the behavior the embedded representation of item i under, e ik is the embedded representation of item i under behavior k, e jk is the embedded representation of item j under behavior k; K is the total number of behaviors, τ is the set hyperparameter, represents the set of users, represents the set of items, and T represents the transpose.

[0033] In step 2.4 above, when performing the second self-supervised contrastive learning: First, two different user noises are respectively added to the embedded representations of users under each behavior to obtain the enhanced embedded representations of two users under each behavior. At the same time, two different item noises are respectively added to the embedded representations of items under each behavior to obtain the enhanced embedded representations of two items under each behavior; then, the user contrast view pairs are constructed using the enhanced embedded representations of two users under each behavior to achieve the self-supervised contrastive learning of users. At the same time, the item contrast view pairs are constructed using the enhanced embedded representations of two items under each behavior to achieve the self-supervised contrastive learning of items.

[0034] The loss of the second self-supervised contrastive learning task

[0035]

[0036] Among them, is the user positive pair of the user comparison view pair in the second self-supervised contrastive learning task, is the user negative pair of the user comparison view pair in the second self-supervised contrastive learning task, e′ uk is the first enhanced embedding representation of user u under behavior k, e″ uk is the second enhanced embedding representation of user u under behavior k, e″ vk is the second enhanced embedding representation of user v under behavior k; is the item positive pair of the item comparison view pair in the second self-supervised contrastive learning task, is the item negative pair of the item comparison view pair in the second self-supervised contrastive learning task; e′ ik T is the first enhanced embedding representation of item i under behavior k, e″ ik is the second enhanced embedding representation of item i under behavior k, e″ jk is the second enhanced embedding representation of item j under behavior k; K is the total number of behaviors, and τ is a set hyperparameter, represents the set of users, represents the set of items, and T represents transpose.

[0037] Compared with the prior art, the present invention has the following characteristics:

[0038] 1. Compared with the single-behavior recommendation algorithm, the multi-behavior recommendation algorithm of the present invention considers the multi-behavior historical interaction information of users, and uses supervised learning and self-supervised learning to improve the recommendation performance of the graph neural network that fuses the multi-behaviors of users;

[0039] 2. In the supervised learning of the graph neural network that fuses the multi-behaviors of users, the method of graph convolution is used to learn the multi-behavior historical interaction information of users, capture the high-order connectivity and structural connectivity between users and items, and obtain high-quality embedding representations of users and items;

[0040] 3. In the self-supervised learning of the graph neural network that fuses the multi-behaviors of users, one self-supervised learning task is to perform contrastive learning between the target behavior and other auxiliary behaviors to capture the relationship between the target behavior and the auxiliary behaviors, and the other self-supervised learning task is the contrastive learning task for a single behavior to improve the uniformity of the multi-behavior representation and alleviate the sparsity of the data; the self-supervised learning helps the representation learn more information and improves the generalization ability and recommendation accuracy of the graph neural network that fuses the multi-behaviors of users;

[0041] 4. Traditional recommendation models only consider the accuracy of recommendations, while the recommendation system of the present invention also considers the novelty of recommendations. That is, by controlling the normalization coefficient in the graph convolution process, it can ensure that the recommendation system improves the novelty of recommendations without sacrificing the accuracy of recommendations and alleviates the popularity bias. Description of the Drawings

[0042] Figure 1 It is a flowchart of a graph neural network recommendation method that fuses multiple user behaviors.

[0043] Figure 2 It is a model framework diagram of the present invention. Detailed Embodiment

[0044] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific examples.

[0045] Refer to Figure 1 and Figure 2 , a graph neural network recommendation method that fuses multiple user behaviors, includes the following steps:

[0046] Step 1: Download publicly available multi-behavior data from the Internet, preprocess the multi-behavior data to obtain a multi-behavior data set, and divide the multi-behavior data into a multi-behavior training set and a multi-behavior test set.

[0047] In this example, the multi-behavior data mainly includes historical behavior data of users on items (such as browsing, favoriting, adding to cart, and / or purchasing behaviors) obtained from existing websites (such as Taobao, JD.com, and / or Pinduoduo, etc.). The multi-behavior data is preprocessed to filter users with too many and too few interactions with items, re-number the users and items in the data, and calculate the number of times users interact with items for each behavior, so as to achieve the purpose of data preprocessing.

[0048] Step 2: Use the multi-behavior training set to train the graph neural network that fuses multiple user behaviors, and optimize the comprehensive loss of the trained graph neural network that fuses multiple user behaviors by adjusting all trainable parameters of the trained graph neural network that fuses multiple user behaviors, so as to obtain the trained graph neural network that fuses multiple user behaviors.

[0049] Step 2.1: Use the graph neural network that fuses multiple user behaviors to obtain the embedding representation e of user u under each behavior k uk and the embedding representation e of item i ik , and merge the embedding representation e of user u under all behaviors k uk and the embedding representation e of item i ik to obtain the final representation e of the user uand the final representation of the project. k = 1, 2,..., K, where K is the total number of behaviors; u = 1, 2,..., M, where M is the total number of users; i = 1, 2,..., N, where N is the total number of projects.

[0050] The graph neural network that fuses multiple user behaviors first uses the graph convolution method to capture the high-order connectivity and structural connectivity between users and projects, and obtains the final representations of users and projects, which mainly include three parts: neighbor aggregation, layer combination, and merging:

[0051] 1) Neighbor aggregation

[0052] Neighbor aggregation uses a simplified weighted sum to iteratively calculate the embedding representation of user u under each behavior k at each layer and the embedding representation of project i The specific formula is as follows:

[0053]

[0054] Among them, is the embedding representation of user u under behavior k at the l-th layer, is the embedding representation of user u under behavior k at the (l - 1)-th layer, is the embedding representation of project i under behavior k at the l-th layer, is the embedding representation of project i under behavior k at the (l - 1)-th layer, is the set of projects that interact with user u, is the set of users that interact with project i.

[0055] 2) Layer combination

[0056] The layer combination of graph convolution uses average pooling to combine the embedding representations of user u under each behavior k at each layer and the embedding representation of project i to obtain the embedding representation e of user u under each behavior k uk and that of project i

[0057] embedding representation e ik . The specific formula is as follows:

[0058]

[0059] Among them, e uk is the embedding representation of user u under behavior k after graph convolution, and e ik is the embedding representation of project i under behavior k after graph convolution, and L is the number of layers of graph convolution.

[0060] 3) Merging

[0061] For the embedding representation e of user u under each behavior k uk, consider the number and impact intensity of different types of behaviors to distinguish the intensity of multiple behaviors, and use the fusion operation to combine the embedded representations e of user u under each behavior k uk to obtain the final representation e of user u u . The formula for the fusion operation is:

[0062]

[0063] where e uk is the embedded representation of user u under behavior k, a uk is the semantic fusion coefficient of the k-th behavior of user u, W and b are the weights and biases in the fully connected layer, and σ(■) represents the Relu activation function.

[0064] Semantic fusion is set for different behavior representations of different users. Taking the semantic fusion coefficient a uk of the k-th behavior of user u as an example, it not only needs to consider the proportion of the k-th behavior of user u among all behaviors, but also needs to identify the intensity of different behaviors for all users. Therefore, the semantic fusion coefficient a uk of the k-th behavior of user u is:

[0065]

[0066] where w k is the intensity weight of behavior k, which is the same for all users and has the ability of automatic learning in the model; n uk is the number of interactions that user u has under behavior k, and K is the total number of behaviors.

[0067] For the embedded representation e ik of item i under each behavior k, since the features of the user are static, the direct concatenation operation combines the embedded representations e ik of item i under each behavior k to obtain the final representation e i of item i. The formula for the concatenation operation is:

[0068] e i = g{Cat(e ik )} (5)

[0069] where e ik is the embedded representation of item i under behavior k, Cat(■) represents the concatenation operation between k vectors, and g(■) represents the multi-layer perceptron (MLP).

[0070] In this embodiment, the graph neural network that fuses multiple user behaviors adopts a three-layer LightGCN.

[0071] Step 2.2: Perform supervised learning on the graph neural network that fuses multiple user behaviors, and calculate the loss of the supervised learning task.

[0072] Use the ranking loss BPR loss to calculate the loss of the supervised learning task

[0073]

[0074] where \(O =\{(u, i, j)|(u, i)\in O + ,(u, j)\in O -}\) is the multi-behavior training set, \(O + represents the set of users and items that interact in the multi-behavior training set, \(O - represents the set of users and items that do not interact in the multi-behavior training set, \(e i is the final representation of item \(i\) that has an interaction with user \(u\), \(e j is the final representation of item \(j\) that has no interaction with user \(u\), \(e u is the final representation of user \(u\), \(T\) represents transpose, and \(\sigma(\cdot)\) represents the Relu activation function.

[0075] Step 2.3: Perform self-supervised learning on the graph neural network that fuses multiple user behaviors, and calculate the loss of the self-supervised learning task.

[0076] Through the self-supervised learning task, help the representation capture more information and further improve the recommendation performance. The self-supervised learning includes 2 contrastive learning tasks. In contrastive learning, the views of the contrastive view pairs for the same node are regarded as positive pairs, and the views of the contrastive view pairs for any different nodes are regarded as negative pairs. After finding the positive pairs and negative pairs, the core idea is to maximize the consistency of the positive pairs between different views, that is, to maximize the consistency of the positive pairs and minimize the consistency of the negative pairs.

[0077] 1) The first self-supervised contrastive learning task is to perform contrastive learning between the target behavior and each auxiliary behavior, which can capture the commonalities of multiple behaviors.

[0078] Based on the characteristics of the training data set, specify one of the \(K\) behaviors as the target behavior, then the other \(K - 1\) behaviors are used as auxiliary behaviors, and contrastive view pairs are constructed using the target behavior and other auxiliary behaviors. For example, if the training data set uses an e-commerce data set with behaviors such as purchase, browsing, and adding to the shopping cart, then purchase is specified as the target behavior, and behaviors such as browsing and adding to the shopping cart are specified as auxiliary behaviors.

[0079] Use the contrastive loss InfoNCE loss to calculate the loss of the first self-supervised contrastive learning task Maximize the consistency of the positive pairs and minimize the consistency of the negative pairs, that is:

[0080]

[0081] is the user contrast learning loss function in the first self-supervised contrast learning task:

[0082]

[0083] is the item contrast learning loss function in the first self-supervised contrast learning task:

[0084]

[0085] where τ is the set hyperparameter, T represents the transpose, represents the target behavior, and k represents the auxiliary behavior, represents the set of users, and J represents the set of items, is the positive user pair in the first self-supervised contrast learning task, is the negative user pair in the first self-supervised contrast learning task, is the positive item pair in the first self-supervised contrast learning task, is the negative item pair in the first self-supervised contrast learning task.

[0086] 2) The second self-supervised contrast learning task is to perform contrast learning on a single behavior, which can improve the uniformity of multi-behavior representations and alleviate the sparsity of data.

[0087] Referring to the idea of SimGCL, noise is added to the direct embedding representation to obtain an enhanced view, replacing the method of using data augmentation in traditional contrast learning models to generate enhanced views, and using the enhanced views to construct contrast view pairs.

[0088] First, user noise Δ′ uk and Δ″ u are added to the embedding representation e u of user u under behavior k respectively, uk to obtain two enhanced embedding representations e′ uk and e″ uk of user u under behavior k; then these two enhanced embedding representations e′ uk and e″

[0089] e′ uk = e uk + Δ′ u (10)

[0090] e″ uk = e ik + Δ″ u(11)

[0091] First, add item noise Δ′ ik and Δ″ i to the embedding representation e of item i under behavior k i to obtain two enhanced embedding representations e′ ik and e″ ik of item i under behavior k; then use these two enhanced embedding representations e′ ik and e″ ik of item i under behavior k to construct a contrastive view pair of the item.

[0092] e′ ik = e ik + Δ′ i (12)

[0093] e″ ik = e ik + Δ″ i (13)

[0094] where Δ′ u , Δ″ u are the added user noise vectors, and Δ′ i , Δ″ i are the added item noise vectors. The user noise vectors Δ′ u , Δ″ u and the item noise vectors Δ′ i , Δ″ i both satisfy the constraint condition of random uniform noise Δ ~ U(0, 1).

[0095] Adopt the contrastive loss InfoNCE loss to calculate the loss of the second self-supervised contrastive learning task to maximize the consistency of positive pairs and minimize the consistency of negative pairs, that is:

[0096]

[0097] is the user contrastive learning loss function in the second self-supervised contrastive learning task:

[0098]

[0099] is the item contrastive learning loss function in the second self-supervised contrastive learning task:

[0100]

[0101] where τ is the set hyperparameter, T represents the transpose, represents the set of users, Represents a set of items, Is the user positive pair for the second self-supervised contrastive learning task, Is the user negative pair for the second self-supervised contrastive learning task, Is the item positive pair for the second self-supervised contrastive learning task, Is the item negative pair for the second self-supervised contrastive learning task.

[0102] Step 2.4: Combine the loss of supervised learning in Step 2.2 and the loss of self-supervised learning in Step 2.3, and optimize the comprehensive loss of the graph neural network that fuses multi-behavior of users by adjusting the trainable parameters of the graph neural network that fuses multi-behavior of users Obtain the trained graph neural network that fuses multi-behavior of users;

[0103]

[0104] Among them, Is the supervised learning loss, Is the first self-supervised contrastive learning task loss, Is the second self-supervised contrastive learning task loss, λ1 is the weight of the first self-supervised contrastive learning task loss, λ2 is the weight of the second self-supervised contrastive learning task loss, Θ is all trainable parameters in the supervised learning and self-supervised learning tasks, Represents L2 regularization, and μ is the set hyperparameter of L2 regularization.

[0105] Use the gradient descent optimization algorithm to optimize the trainable parameters of the graph neural network that fuses multi-behavior of users. In this example, the gradient descent optimization algorithm adopted is the adam algorithm (adaptive motion estimation algorithm).

[0106] Step 3: Use the multi-behavior test set to test the trained graph neural network that fuses multi-behavior of users, and optimize the evaluation metrics of the accuracy and novelty of the trained graph neural network that fuses multi-behavior of users by adjusting the graph convolution normalization coefficient of the trained graph neural network that fuses multi-behavior of users, and obtain the final graph neural network that fuses multi-behavior of users.

[0107] In the graph neural network that fuses multi-behavior of users, the actual calculation of neighbor aggregation is in matrix form as follows:

[0108] E k (l) =(D -r A k D -(1-r) )E k (l-1) (18)

[0109] Among them, E k(0) is randomly initialized to size The embedding representation of users and items under the behavior k of the 0th layer; E k (l-1) Is the size of The embedding representation of users and items under the behavior k of the l-1th layer of The embedding representation of user u under the behavior k of the l-1th layer and size The embedding representation of item i under row k of the l-1th layer Composition; d is the dimension, M is the total number of users, N is the total number of projects; A k is the adjacency matrix, R k is the interaction matrix R between the behavior k user and the item k , the interaction matrix R k The element in row i and column j of indicates whether user i and item j have interacted under behavior k: when user i and item j have interacted under behavior k, then otherwise, D k It is a (M+N)×(M+N) diagonal matrix; r is the graph convolution normalization coefficient.

[0110] In common graph convolution recommendation algorithms, the graph convolution normalization coefficient r is generally 0.5. When r<1, the recommendation model tends to recommend nodes with high degrees (hot items), and the accuracy of the recommendation model is improved but the novelty is reduced; when r>1, the recommendation model tends to recommend nodes with low degrees (unpopular items), and the novelty of the recommendation model is improved but the accuracy is reduced. To this end, the graph convolution normalization coefficient r of the present invention is set in the range of [0.5, 0.55, ..., 1.5] to repeat the training process of step 2 multiple times, and select the optimal result based on the evaluation indicators of the accuracy and novelty of the recommendation model.

[0111] The accuracy evaluation indicator Recall@Q is:

[0112]

[0113] The novelty evaluation index Nov@Q is:

[0114]

[0115] Where Q is the set value, indicating the first Q recommended items. R(u) represents the set of recommended items predicted by the model, and T(u) represents the set of recommended items in the real multi-behavior test set. The Recall@Q of the entire dataset can be obtained by averaging the Recall@Q obtained for each user. u represents the set of users, I u (Q) represents the Top-Q recommended item set of user u, d i It indicates the number of times item i is observed in the multi-behavior training set, that is, the degree of node i. According to the definition, the lower the value of Nov@Q, the more popular the top-ranked items are. When the value of Recall@Q is large, it means that the recommendation system can recommend more items that users are really interested in to them.

[0116] The graph convolution normalization system of the graph neural network that integrates multiple user behaviors is optimized using the gradient descent optimization algorithm. In this example, the gradient descent optimization algorithm used is the Adam algorithm (adaptive motion estimation algorithm).

[0117] Step 4: Use the final graph neural network that integrates multiple user behaviors to predict the recommended items for the current user, and recommend the predicted results to the current user.

[0118] The final graph neural network that integrates multiple user behaviors uses the user’s final representation u and the final representation of the project i Perform inner product as recommendation ranking score

[0119]

[0120] in, The ranking score of item i for user u is predicted by the final graph neural network that integrates multiple user behaviors. By calculating the ranking scores of all items for the current user, the item with the largest ranking score (TOP-Q) can be selected as the final recommended item for the current user.

[0121] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. A graph neural network recommendation method that integrates multiple user behaviors, characterized in that, The steps are as follows: Step 1: Download publicly available multi-behavior data from the Internet, preprocess the multi-behavior data to obtain a multi-behavior data set, and divide the multi-behavior data into a multi-behavior training set and a multi-behavior test set; Step 2: Use the multi-behavior training set to train the graph neural network that fuses users' multi-behaviors. Optimize the comprehensive loss of the trained graph neural network that fuses users' multi-behaviors by adjusting all trainable parameters of the trained graph neural network that fuses users' multi-behaviors to obtain the trained graph neural network that fuses users' multi-behaviors, that is: Step 2.1: Use the graph neural network that fuses users' multi-behaviors to obtain the embedding representations of users and items under each behavior, and merge the embedding representations of users and items under all behaviors to obtain the final representations of users and items; Step 2.2: Conduct supervised learning on the graph neural network that fuses users' multi-behaviors, and calculate the loss of the supervised learning task using ranking loss; Step 2.3: Conduct the first self-supervised contrast learning on the graph neural network that fuses users' multi-behaviors, that is, conduct contrast learning between the target behavior and each auxiliary behavior, and calculate the loss of the first self-supervised contrast learning task using contrast loss; Step 2.4: Conduct the second self-supervised contrast learning on the graph neural network that fuses users' multi-behaviors, that is, conduct contrast learning for a single behavior, and calculate the loss of the second self-supervised contrast learning task using contrast loss; Step 2.

5. Combine the losses in Steps 2.2 - 2.4, and optimize the comprehensive loss of the graph neural network that fuses multiple user behaviors by adjusting the trainable parameters of the graph neural network that fuses multiple user behaviors, so as to obtain the trained graph neural network that fuses multiple user behaviors; where the comprehensive loss is as follows: Among them, is the loss of the supervised learning task, is the loss of the first self-supervised contrastive learning task, is the loss of the second self-supervised contrastive learning task. λ1 is the weight of the loss of the first self-supervised contrastive learning task, λ2 is the weight of the loss of the second self-supervised contrastive learning task, Θ are all trainable parameters of the graph neural network that fuses multiple user behaviors, and μ is the hyperparameter of L2 regularization, represents L2 regularization; Step 3: Use the multi-behavior test set to test the trained graph neural network that fuses users' multi-behaviors. Optimize the evaluation metrics of the accuracy and novelty of the trained graph neural network that fuses users' multi-behaviors by adjusting the graph convolution normalization coefficient of the trained graph neural network that fuses users' multi-behaviors to obtain the final graph neural network that fuses users' multi-behaviors; Step 4: Use the final graph neural network that fuses users' multi-behaviors to predict the recommended items for the current user, and recommend the prediction results to the current user.

2. The graph neural network recommendation method integrating multiple user behaviors according to claim 1, characterized in that In step 2.1, the embedding representations of the user under all behaviors are merged using a fusion operation to obtain the final representation of the user; the final representation \(e\) of user \(u\) is: u is: where, e uk is the embedded representation of user u under behavior k, a uk is the semantic fusion coefficient of the k-th behavior of user u, K is the total number of behaviors, σ(■) represents the Relu activation function, and W and b are the weights and biases in the fully connected layer of the graph neural network that fuses the multiple behaviors of the user.

3. The graph neural network recommendation method integrating multiple user behaviors according to claim 2, characterized in that, The semantic fusion coefficient a of the k-th behavior of user u uk is as follows: where, w k is the intensity weight of behavior k, n uk is the number of interactions that user u has under behavior k, and K is the total number of behaviors.

4. A graph neural network recommendation method that integrates multiple user behaviors according to claim 1, characterized in that In step 2.1, the embedding representations of items under all behaviors are merged using a cascading operation to obtain the final representation of the item; the final representation e of item i i is as follows: e i = g{Cat(e ik )} where, e ik is the embedding representation of item i under each behavior k, Cat(■) represents the concatenation operation of the graph neural network that fuses multiple user behaviors, and g(■) represents the multi-layer perception operation of the graph neural network that fuses multiple user behaviors.

5. The graph neural network recommendation method integrating multiple user behaviors according to claim 1, characterized in that, In Step 2.2, the loss of the supervised learning task is as follows: where \(O =\{(u, i, j)\mid(u, i)\in O + ,(u, j)\in O - \}\) is the multi-behavior training set, \(O + represents the set of users and items that interact in the multi-behavior training set, \(O - represents the set of users and items that do not interact in the multi-behavior training set, \(e i is the final representation of item \(i\), \(e j is the final representation of item \(j\), \(e u is the final representation of user \(u\), \(T\) represents transpose, and \(\sigma(\cdot)\) represents the Relu activation function.

6. The graph neural network recommendation method integrating multiple user behaviors according to claim 1, characterized in that In Step 2.3, when conducting the first self-supervised contrast learning: First, based on the characteristics of the multi-behavior training set, designate one of all behaviors as the target behavior, and the remaining behaviors as auxiliary behaviors; Then, use the embedding representations of users under the target behavior and the embedding representations of users under other auxiliary behaviors to construct user contrast view pairs to achieve self-supervised contrast learning of users, and at the same time use the embedding representations of items under the target behavior and the embedding representations of items under other auxiliary behaviors to construct item contrast view pairs to achieve self-supervised contrast learning of items.

7. A graph neural network recommendation method that integrates multiple user behaviors according to claim 1 or 6, characterized in that, In step 2.3, the loss of the first self-supervised contrastive learning task is as follows: Among them, is the positive user pair of the user comparison view pair in the first self-supervised contrastive learning task, is the negative user pair of the user comparison view pair in the first self-supervised contrastive learning task, is the behavior the embedded representation of user u under, e uk is the embedded representation of user u under behavior k, e vk is the embedded representation of user v under behavior k; is the positive item pair of the item comparison view pair in the first self-supervised contrastive learning task, is the negative item pair of the item comparison view pair in the first self-supervised contrastive learning task, is the behavior the embedded representation of item i under, e ik is the embedded representation of item i under behavior k, e jk is the embedded representation of item j under behavior k; K is the total number of behaviors, and τ is a set hyperparameter, represents the set of users, represents the set of items, and T represents the transpose.

8. A graph neural network recommendation method integrating multiple user behaviors according to claim 1, characterized in that, In step 2.4, when performing the second self-supervised contrastive learning: first, add two different user noises to the embedded representations of users under each behavior respectively to obtain the enhanced embedded representations of two users under each behavior, and at the same time, add two different item noises to the embedded representations of items under each behavior respectively to obtain the enhanced embedded representations of two items under each behavior; then, use the enhanced embedded representations of two users under each behavior to construct user contrastive view pairs to achieve self-supervised contrastive learning of users, and at the same time, use the enhanced embedded representations of two items under each behavior to construct item contrastive view pairs to achieve self-supervised contrastive learning of items.

9. The graph neural network recommendation method integrating multiple user behaviors according to claim 1 or 8, characterized in that, The loss of the second self-supervised contrastive learning task Among them, is the user positive pair of the user comparison view pair in the second self-supervised contrastive learning task, is the user negative pair of the user comparison view pair in the second self-supervised contrastive learning task, e′ uk is the first enhanced embedding representation of user u under behavior k, e″ uk is the second enhanced embedding representation of user u under behavior k, e″ vk is the second enhanced embedding representation of user v under behavior k; is the item positive pair of the item comparison view pair in the second self-supervised contrastive learning task, is the item negative pair of the item comparison view pair in the second self-supervised contrastive learning task; e′ ik T is the first enhanced embedding representation of item i under behavior k, e″ ik is the second enhanced embedding representation of item i under behavior k, e″ jk is the second enhanced embedding representation of item j under behavior k; K is the total number of behaviors, and τ is the set hyperparameter, represents the set of users, represents the set of items, and T represents the transpose.

Images