Target behavior prediction method based on neighbor context self-supervised contrast learning

By introducing a method based on neighbor context self-supervised comparison learning in a multi-behavior recommendation system, using graph neural network and self-attention mechanism to optimize node representations of users and products, the problem of existing systems ignoring neighbor collaborative information is solved, and the problem of existing systems ignoring neighbor collaborative information is improved Accuracy and recommended performance of target behavior prediction.

CN119067753APending Publication Date: 2024-12-03SHANXI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411184962.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing multi-behavior recommendation system ignores the collaborative behavior information of neighbor positive samples, resulting in inaccurate user interest expression and affecting the accuracy of target behavior prediction.

Method used

Using a method based on neighbor context self-supervised contrast learning, the node representation of users and products is learned through graph neural network, and a self-attention mechanism is used to model multiple behavior dependencies, adaptively select neighbor positive samples, calculate self-supervised contrast learning losses, and optimize User and product node representations and optimize click interest scores through non-sampling strategies.

Benefits of technology

Improve the accuracy of target behavior prediction and the performance of recommendation system, generate more accurate product recommendation lists, and enhance personalized service recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067753A_ABST
    Figure CN119067753A_ABST
Patent Text Reader

Abstract

The invention discloses a target behavior prediction method based on neighbor context self-supervised contrast learning, and belongs to the technical field of service recommendation. Aiming at the fact that an existing system neglects the effect of collaborative behavior information of neighbor positive samples in learning user interest representation, firstly, a graph neural network and a self-attention mechanism are utilized to model a dependency relationship among different behaviors of a user, and user and product node representation is enhanced; a neighbor positive sample of a target user is adaptively selected, a self-supervised contrast learning strategy based on neighbor perception is designed, the semantic similarity between the neighbor positive sample of a similar target behavior and the target user is increased, and user and product node representation is optimized; and calculating a target behavior prediction score of the user to the product based on the optimized user and product node representation, and generating a product recommendation list. The accuracy of target behavior prediction of a recommendation system is improved by learning behavior interests of neighbor users and personalized weights among multiple behaviors of target users, and main application is provided in the aspect of personalized recommendation systems.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of service recommendation, and in particular relates to a target behavior prediction method based on neighbor context self-supervised contrastive learning. Background Art

[0002] The multi-behavior recommendation system mainly learns the user's behavioral interest preferences by analyzing the user's various auxiliary behaviors (such as clicking, browsing, and collecting, etc.), and then predicts the user's target behavior (such as purchasing) and recommends products that the user may be interested in. To a certain extent, it can accurately predict the user's behavioral intentions and improve the click-through conversion rate of the commercial platform. At present, most multi-behavior recommendation systems fuse the representations of users under different behaviors to obtain enhanced representations of users and realize target behavior prediction based on enhanced representations. However, there are complex dependencies between multiple behaviors, and different users have different multi-behavior patterns. How to discover similar neighbor users based on multiple behaviors is of great significance to improving the accuracy of target behavior prediction.

[0003] In a multi-behavior recommendation system, understanding the behavioral dependency between the target behavior and auxiliary behaviors is an important part of multi-behavior recommendation. Based on graph neural networks, we capture the dependency between different behaviors and learn multi-behavior-aware user and product representations. Then, by finding neighbor samples with similar multi-behaviors to the target user, we use the comparative learning strategy of neighbor positive samples and target users to capture the neighbors' collaborative behavior interest information. At the same time, we consider the comparative similarity of node representations under different views of users, and use multi-task joint learning to enhance the representation quality of nodes, improve the accuracy of target behavior prediction, and improve the product recommendation performance of the multi-behavior recommendation system. Summary of the invention

[0004] In order to solve the problem that the existing multi-behavior recommendation system ignores the role of collaborative behavior information of neighbor positive samples in learning user interest representation, the present invention provides a target behavior prediction method based on neighbor context self-supervised contrastive learning.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The target behavior prediction method based on neighbor context self-supervised contrastive learning comprises the following steps:

[0007] Step S1, using graph neural network to learn node representations of users and products under single behavior respectively;

[0008] Furthermore, the specific steps of step S1 are:

[0009] Step 1.1: According to different types of interaction behaviors between users and products, extract the interaction information between users and products under different single behaviors, and construct the user-product interaction subgraph G under single behavior k. k ; User-product interaction subgraph G under single behavior k k The graph convolution operation is performed using the graph neural network to learn the embedding propagation of nodes. The user node is represented as:

[0010]

[0011] In formula (1), LeakyReLu is a nonlinear activation function. represents the embedding representation of user u at the l+1th propagation layer under single behavior k, represents the embedding representation of product i in the lth propagation layer under single behavior k, represents the number of products that user u interacts with in a single behavior k, represents the number of users associated with product i under single behavior k, ⊙ represents the element-by-element product, represents the learnable parameters for users belonging to a single behavior k, is the embedding representation of a single row as the k-th propagation layer;

[0012] Step 1.2, the propagation process of the embedding representation of the lth propagation layer with a single row k to the l+1th layer in the graph neural network is:

[0013] (2)

[0014] In formula (2), is the learnable parameter of single behavior k in the lth propagation layer; d is the embedding representation dimension of users and products;

[0015] Step 1.3, repeat step 1.2, average pool the embedding representation of each layer under single behavior k, and obtain the embedding representation of single behavior k as:

[0016]

[0017] In the formula, mean represents average pooling, represents the summation, and L is the number of propagation layers.

[0018] Step S2, using the self-attention mechanism to model the dependencies between multiple user behaviors and learn node representations for multi-behavior perception;

[0019] Furthermore, the specific steps of step S2 are:

[0020] Step 2.1: fuse the representations of the lth propagation layer of user u in different single behaviors, and denote the multi-behavior of user u in the lth propagation layer as The self-attention mechanism is used to calculate the representation of a single behavior under k The dependency relationship between and multiple behavior representations is defined as:

[0021]

[0022] In formula (4), is the transpose of the multi-behavior representation of the lth propagation layer of user u; d is the embedding representation dimension of users and products;

[0023] Step 2.2, based on the dependency in formula (4), the node representation of multi-behavior perception of user u in the lth propagation layer under single behavior k is obtained as:

[0024]

[0025] Step 2.3, average pool the node representation of multi-behavior perception of each layer of user u under single behavior k, and obtain the node representation of multi-behavior perception of user u under single behavior k as follows:

[0026]

[0027] In the formula, mean represents average pooling, It means sum;

[0028] Step 2.4, repeat steps 2.1-2.3, average pool the node representation of multi-behavior perception of each layer of product i under single behavior k, and obtain the node representation of multi-behavior perception of product i under single behavior k as follows:

[0029]

[0030] In formula (7), It is the node representation of multi-behavior perception of product i under single behavior k in the Lth propagation layer.

[0031] Step S3, adaptively selecting the user's neighbor positive samples under the target behavior, and calculating the neighbor context-aware self-supervised contrastive learning loss to optimize the representation of user and product nodes;

[0032] Furthermore, the specific steps of step S3 are:

[0033] Step 3.1: record the set of commonly clicked products of users u and u′ under target behavior K as are the clicked product sets of user u and u' under target behavior K respectively; in the set U of all users, find all users with the maximum number of interaction records of the same product under target behavior K for user u to form a neighbor positive sample set, recorded as Defined as:

[0034]

[0035] In formula (8), u * represents a positive example of a neighbor of user u, For users u and u * The number of commonly clicked products under target behavior K; is the maximum number of products clicked by user u and any user u′ under target behavior K, where U is the set of all users;

[0036] Step 3.2, in the neighbor positive sample set , calculate u * Similarity of interest with target user u It is used to measure the importance of different neighbor embedding representations and is defined as:

[0037]

[0038] In formula (9), is the set of positive samples of neighbors of user u;

[0039] Step 3.3: Weight all neighbor positive samples and calculate the collaborative embedding of user u’s neighbor positive samples under single behavior k:

[0040]

[0041] In formula (10), Represents user u * Multi-behavior-aware node representation under single behavior k;

[0042] Step 3.4: For user u, embed the neighbor positive samples under single behavior k into a co-embedded representation and user embedding representation is regarded as a positive example, and other users’ embedding representations are regarded as negative examples. The self-supervised contrastive learning loss based on neighbor positive samples is calculated as:

[0043]

[0044] In formula (11), Remove the positive samples of user u's neighbors from the set U of all users A collection of; for and The cosine similarity between ; for and The cosine similarity between ; τ represents the hyperparameter;

[0045] Step 3.5: Based on the embedding representation of user u under the target behavior K, supervise the embedding representation of other single behaviors k. The self-supervised contrastive learning loss is:

[0046]

[0047] In formula (12), for and The cosine similarity between for and The cosine similarity between is the node representation of multi-behavior perception of user u under target behavior K, is the node representation of user u’s multi-behavior perception under other single behaviors k, is the node representation of multi-behavior perception of user v under other single behaviors k;

[0048] Step 3.6, under other K-1 different types of single behaviors k, calculate the self-supervised contrastive learning loss of all users as:

[0049]

[0050] Step 3.7, repeat step 3.5, and calculate the self-supervised contrastive learning loss for all products as:

[0051]

[0052] In formula (14), I is the set of all products, I\{i} is the set of I without product i, is the embedding representation of product i under target behavior K, are the node representations of multi-behavior perception of product i and i' under single behavior k;

[0053] In step 3.8, the sum of the self-supervised contrastive learning losses of users and products is:

[0054]

[0055] Step S4, calculating the multi-behavior click interest scores of users for the product under different behaviors, optimizing the multi-behavior click interest scores based on the non-sampling strategy, and updating the model weight parameters in conjunction with the self-supervised contrastive learning loss;

[0056] Furthermore, the specific steps of step S4 are:

[0057] Step 4.1, based on the embedding representation dimension d of user u and product i, calculate the click interest score of user u and product i under single behavior k:

[0058]

[0059] In formula (16), n is the loop counter of the embedding dimension of users and products; is the value of the nth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (n) is the value of the nth dimension of the embedding representation of a single row k; is the value of the nth dimension of the node representation of multi-behavior perception of product i under single behavior k;

[0060] Step 4.2: Optimize the click interest score based on the non-sampling strategy. For a user set U consisting of a batch of users b And the product set I consisting of all commodities, the non-sampling loss is defined as:

[0061]

[0062] In formula (17), is the set of products that user u interacts with in single behavior k, is the click interest score of user u and product i under single behavior k, Click on the interest score item The positive example weight and negative example weight of m and n are the embedding dimension loop counters of users and products; is the value of the mth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the value of the mth dimension of the node representation of multi-behavior perception of product i under single behavior k;

[0063] Step 4.3, by performing a weighted summation of the non-sampled losses under each single line k, defined as:

[0064]

[0065] In formula (18), λ k is the weighting coefficient of non-sampling loss;

[0066] Step 4.4, combine the unsampled loss and the self-supervised contrastive learning loss through the hyperparameter μ, defined as:

[0067]

[0068] Step S5, based on the learned user and product node representations, calculate the user's interest score for the new product with respect to the target behavior, sort the products according to the scores of the target behavior, and generate a product recommendation list;

[0069] Furthermore, the specific steps of step S5 are:

[0070] Embedding representation r based on target behavior K K , Node representation of user's multi-behavior perception under target behavior K And the node representation of multi-behavior perception of new product i' under target behavior K Calculate the click interest score of user u and new product i' in target behavior K as:

[0071]

[0072] In formula (20), is the value of the nth dimension of the node representation of the multi-behavior perception of user u under the target behavior K; r K (n) is the value of the nth dimension of the embedding representation of the target behavior K; is the value of the nth dimension of the node representation of multi-behavior perception of the new product i' under the target behavior K;

[0073] For multiple new products, calculate the predicted scores of users' interest in different products, and recommend the top t products with higher predicted scores to users.

[0074] Compared with the prior art, the present invention has the following advantages:

[0075] The method provided in the present invention is different from the existing methods in that it designs a multi-behavior recommendation method based on neighbor context self-supervised comparative learning, fully utilizes the collaborative signals of neighbor users with similar target behaviors, constructs a target user behavior prediction model based on neighbor context information, realizes multi-behavior prediction based on neighbor context self-supervised comparative learning, and generates a product recommendation list based on the click interest of the target behavior. The present invention enhances the behavior representation characteristics of the target user and improves the accuracy of the target user behavior prediction recommendation by introducing the user's neighbor context behavior preference information learning and the behavior similarity between the target user, and can provide major applications in the multi-behavior personalized service recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a schematic diagram of the overall model architecture of the present invention. DETAILED DESCRIPTION

[0077] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0078] The target behavior prediction method based on neighbor context self-supervised contrastive learning of the present invention is implemented by a computer program. The specific implementation of the technical solution proposed by the present invention will be described in detail below according to the process.

[0079] like Figure 1 As shown, the target behavior prediction method based on neighbor context self-supervised contrastive learning of the present invention comprises the following steps:

[0080] Step S1, using graph neural network to learn node representations of users and products under single behavior respectively;

[0081] Furthermore, the specific steps of step S1 are:

[0082] Step 1.1: According to different types of interaction behaviors between users and products, extract the interaction information between users and products under different single behaviors, and construct the user-product interaction subgraph G under single behavior k. k ; User-product interaction subgraph G under single behavior k k The graph convolution operation is performed using the graph neural network to learn the embedding propagation of nodes. The user node is represented as:

[0083]

[0084] In formula (1), LeakyReLu is a nonlinear activation function. represents the embedding representation of user u at the l+1th propagation layer under single behavior k, represents the embedding representation of product i in the lth propagation layer under single behavior k, represents the number of products that user u interacts with in a single behavior k, represents the number of users associated with product i under single behavior k, ⊙ represents the element-by-element product, represents the learnable parameters for users belonging to a single behavior k, is the embedding representation of a single row as the k-th propagation layer;

[0085] Step 1.2, the propagation process of the embedding representation of the lth propagation layer with a single row k to the l+1th layer in the graph neural network is:

[0086]

[0087] In formula (2), is the learnable parameter of single behavior k in the lth propagation layer; d is the embedding representation dimension of users and products;

[0088] Step 1.3, repeat step 1.2, average pool the embedding representation of each layer under single behavior k, and obtain the embedding representation of single behavior k as:

[0089]

[0090] In the formula, mean represents average pooling, represents the summation, and L is the number of propagation layers.

[0091] Step S2, using the self-attention mechanism to model the dependencies between multiple user behaviors and learn node representations for multi-behavior perception;

[0092] Furthermore, the specific steps of step S2 are:

[0093] Step 2.1: fuse the representations of the lth propagation layer of user u in different single behaviors, and denote the multi-behavior of user u in the lth propagation layer as The self-attention mechanism is used to calculate the representation of a single behavior under k The dependency relationship between and multiple behavior representations is defined as:

[0094]

[0095] In formula (4), is the transpose of the multi-behavior representation of the lth propagation layer of user u; d is the embedding representation dimension of users and products;

[0096] Step 2.2, based on the dependency in formula (4), the node representation of multi-behavior perception of user u in the lth propagation layer under single behavior k is obtained as:

[0097]

[0098] Step 2.3, average pool the node representation of multi-behavior perception of each layer of user u under single behavior k, and obtain the node representation of multi-behavior perception of user u under single behavior k as follows:

[0099]

[0100] In the formula, mean represents average pooling, It means sum;

[0101] Step 2.4, repeat steps 2.1-2.3, average pool the node representation of multi-behavior perception of each layer of product i under single behavior k, and obtain the node representation of multi-behavior perception of product i under single behavior k as follows:

[0102]

[0103] In formula (7), It is the node representation of multi-behavior perception of product i under single behavior k in the Lth propagation layer.

[0104] Step S3: Adaptively select the user’s neighbor positive samples under the target behavior, calculate the neighbor context-aware self-supervised contrastive learning loss to optimize the representation of user and product nodes

[0105] Furthermore, the specific steps of step S3 are:

[0106] Step 3.1: record the set of commonly clicked products of users u and u′ under target behavior K as are the clicked product sets of user u and u' under target behavior K respectively; in the set U of all users, find all users with the maximum number of interaction records of the same product under target behavior K for user u to form a neighbor positive sample set, recorded as Defined as:

[0107]

[0108] In formula (8), u * represents a positive example of a neighbor of user u, For users u and u * The number of commonly clicked products under target behavior K; is the maximum number of products clicked by user u and any user u′ under target behavior K, where U is the set of all users;

[0109] Step 3.2, in the neighbor positive sample set , calculate u * Similarity of interest with target user u It is used to measure the importance of different neighbor embedding representations and is defined as:

[0110]

[0111] In formula (9), is the set of positive samples of neighbors of user u;

[0112] Step 3.3: Weight all neighbor positive samples and calculate the collaborative embedding of user u’s neighbor positive samples under single behavior k:

[0113]

[0114] In formula (10), Represents user u * Multi-behavior-aware node representation under single behavior k;

[0115] Step 3.4: For user u, embed the neighbor positive samples under single behavior k into a co-embedded representation and user embedding representation is regarded as a positive example, and other users’ embedding representations are regarded as negative examples. The self-supervised contrastive learning loss based on neighbor positive samples is calculated as:

[0116]

[0117] In formula (11), Remove the positive samples of user u's neighbors from the set U of all users A collection of; for and The cosine similarity between ; for and The cosine similarity between ; τ represents the hyperparameter;

[0118] Step 3.5: Based on the embedding representation of user u under the target behavior K, supervise the embedding representation of other single behaviors k. The self-supervised contrastive learning loss is:

[0119]

[0120] In formula (12), for and The cosine similarity between for and The cosine similarity between is the node representation of multi-behavior perception of user u under target behavior K, is the node representation of user u’s multi-behavior perception under other single behaviors k, is the node representation of multi-behavior perception of user v under other single behaviors k;

[0121] Step 3.6, under other K-1 different types of single behaviors k, calculate the self-supervised contrastive learning loss of all users as:

[0122]

[0123] Step 3.7, repeat step 3.5, and calculate the self-supervised contrastive learning loss for all products as:

[0124]

[0125] In formula (14), I is the set of all products, I\{i} is the set of I without product i, is the embedding representation of product i under target behavior K, are the node representations of multi-behavior perception of product i and i' under single behavior k;

[0126] In step 3.8, the sum of the self-supervised contrastive learning losses of users and products is:

[0127]

[0128] Step S4, calculating the multi-behavior click interest scores of users for the product under different behaviors, optimizing the multi-behavior click interest scores based on the non-sampling strategy, and updating the model weight parameters in conjunction with the self-supervised contrastive learning loss;

[0129] Furthermore, the specific steps of step S4 are:

[0130] Step 4.1, based on the embedding representation dimension d of user u and product i, calculate the click interest score of user u and product i under single behavior k:

[0131]

[0132] In formula (16), n is the loop counter of the embedding dimension of users and products; is the value of the nth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (n) is the value of the nth dimension of the embedding representation of a single row k; is the value of the nth dimension of the node representation of multi-behavior perception of product i under single behavior k;

[0133] Step 4.2: Optimize the click interest score based on the non-sampling strategy. For a user set U consisting of a batch of users b And the product set I consisting of all commodities, the non-sampling loss is defined as:

[0134]

[0135] In formula (17), is the set of products that user u interacts with in single behavior k, is the click interest score of user u and product i under single behavior k, Click on the interest score item The positive example weight and negative example weight of m and n are the embedding dimension loop counters of users and products; is the value of the mth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the value of the mth dimension of the node representation of multi-behavior perception of product i under single behavior k;

[0136] Step 4.3, by performing a weighted summation of the non-sampled losses under each single line k, defined as:

[0137]

[0138] In formula (18), λ k is the weighting coefficient of non-sampling loss;

[0139] Step 4.4, combine the unsampled loss and the self-supervised contrastive learning loss through the hyperparameter μ, defined as:

[0140]

[0141] Step S5, based on the learned user and product node representations, calculate the user's interest score for the new product with respect to the target behavior, sort the products according to the scores of the target behavior, and generate a product recommendation list;

[0142] Furthermore, the specific steps of step S5 are:

[0143] Embedding representation r based on target behavior K K , Node representation of user's multi-behavior perception under target behavior K And the node representation of multi-behavior perception of new product i' under target behavior K Calculate the click interest score of user u and new product i' in target behavior K as:

[0144]

[0145] In formula (20), is the value of the nth dimension of the node representation of the multi-behavior perception of user u under the target behavior K; r K (n) is the value of the nth dimension of the embedding representation of the target behavior K; is the value of the nth dimension of the node representation of multi-behavior perception of the new product i' under the target behavior K;

[0146] For multiple new products, calculate the predicted scores of users' interest in different products, and recommend the top t products with higher predicted scores to users.

[0147] In order to verify the effectiveness of the method, experiments were conducted on the Taobao dataset (https: / / tianchi.aliyun.com / dataset / dataDetail?dataId=649), which contains multiple types of user-item interaction behaviors, namely browsing, adding to shopping cart, and purchasing. The dataset information is shown in Table 1:

[0148] Table 1 Dataset

[0149]

[0150] The evaluation indicators used are HR@10 and NDCG@10. In order to verify the effectiveness and advancement of the technical solution proposed in this invention, several existing recommended prediction model methods are selected for comparison: LightGCN, UltraGCN, MB-GMN, EHCF, GHCF and MBSSL. The experimental results are shown in Table 2:

[0151] Table 2 Experimental results

[0152]

[0153] It can be seen from the results in Table 2 that the technical solution of the present invention can obtain detection results with better accuracy and reliability than existing methods when predicting users' interest in project products.

[0154] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A target behavior prediction method based on neighbor context self-supervised contrastive learning, characterized in that: The following steps are involved: Step S1, using graph neural network to learn node representations of users and products under single behavior respectively; Step S2, using the self-attention mechanism to model the dependencies between multiple user behaviors and learn node representations for multi-behavior perception; Step S3, adaptively selecting the user's neighbor positive samples under the target behavior, and calculating the neighbor context-aware self-supervised contrastive learning loss to optimize the representation of user and product nodes; Step S4, calculating the multi-behavior click interest scores of users for the product under different behaviors, optimizing the multi-behavior click interest scores based on the non-sampling strategy, and updating the model weight parameters in conjunction with the self-supervised contrastive learning loss; Step S5, based on the learned user and product node representations, calculate the user's interest score for the new product with respect to the target behavior, sort the products according to the scores of the target behavior, and generate a product recommendation list.

2. The target behavior prediction method based on neighbor context self-supervised contrastive learning according to claim 1 is characterized in that: The specific steps of step S1, using the graph neural network to learn the node representations of users and products under a single behavior are as follows: Step 1.1: According to different types of interaction behaviors between users and products, extract the interaction information between users and products under different single behaviors, and construct the user-product interaction subgraph G under single behavior k. k ; User-product interaction subgraph G under single behavior k k The graph convolution operation is performed using the graph neural network to learn the embedding propagation of nodes. The user node is represented as: In formula (1), LeakyReLu is a nonlinear activation function. represents the embedding representation of user u at the l+1th propagation layer under single behavior k, represents the embedding representation of product i in the lth propagation layer under single behavior k, represents the number of products that user u interacts with in a single behavior k, represents the number of users associated with product i under single behavior k, ⊙ represents the element-by-element product, represents the learnable parameters for users belonging to a single behavior k, is the embedding representation of a single row as the k-th propagation layer; Step 1.2, the propagation process of the embedding representation of the lth propagation layer with a single row k to the l+1th layer in the graph neural network is: In formula (2), is the learnable parameter of single behavior k in the lth propagation layer; d is the embedding representation dimension of users and products; Step 1.3, repeat step 1.2, average pool the embedding representation of each layer under single behavior k, and obtain the embedding representation of single behavior k as: In the formula, mean represents average pooling, represents the summation, and L is the number of propagation layers.

3. The target behavior prediction method based on neighbor context self-supervised contrastive learning according to claim 1 is characterized in that: The specific steps of step S2, using the self-attention mechanism to model the dependency relationship between multiple user behaviors and learning the node representation of multi-behavior perception, are as follows: Step 2.1: fuse the representations of the lth propagation layer of user u in different single behaviors, and denote the multi-behavior of user u in the lth propagation layer as The self-attention mechanism is used to calculate the representation of a single behavior under k The dependency relationship between and multiple behavior representations is defined as: In formula (4), is the transpose of the multi-behavior representation of the lth propagation layer of user u; d is the embedding representation dimension of users and products; Step 2.2, based on the dependency in formula (4), the node representation of multi-behavior perception of user u in the lth propagation layer under single behavior k is obtained as: Step 2.3, average pool the node representation of multi-behavior perception of each layer of user u under single behavior k, and obtain the node representation of multi-behavior perception of user u under single behavior k as follows: In the formula, mean represents average pooling, It means sum; Step 2.4, repeat steps 2.1-2.3, average pool the node representation of multi-behavior perception of each layer of product i under single behavior k, and obtain the node representation of multi-behavior perception of product i under single behavior k as follows: In formula (7), It is the node representation of multi-behavior perception of product i under single behavior k in the Lth propagation layer.

4. The target behavior prediction method based on neighbor context self-supervised contrastive learning according to claim 1 is characterized in that: The specific steps of step S3, adaptively selecting the user's neighbor positive samples under the target behavior, and calculating the neighbor context-aware self-supervised contrastive learning loss to optimize the representation of the user and product nodes are: Step 3.1: record the set of commonly clicked products of users u and u′ under target behavior K as are the clicked product sets of user u and u' under target behavior K respectively; in the set U of all users, find all users with the maximum number of interaction records of the same product under target behavior K for user u to form a neighbor positive sample set, recorded as Defined as: In formula (8), u * represents a positive example of a neighbor of user u, For users u and u * The number of commonly clicked products under target behavior K; is the maximum number of products clicked by user u and any user u′ under target behavior K, where U is the set of all users; Step 3.2, in the neighbor positive sample set , calculate u * Similarity of interest with target user u It is used to measure the importance of different neighbor embedding representations and is defined as: In formula (9), is the set of positive samples of neighbors of user u; Step 3.3: Weight all neighbor positive samples and calculate the collaborative embedding of the neighbor positive samples of user u under single behavior k as follows: In formula (10), Represents user u * Multi-behavior-aware node representation under single behavior k; Step 3.4: For user u, the neighbor positive samples under single behavior k are collaboratively embedded to represent and user embedding representation is regarded as a positive example, and other users’ embedding representations are regarded as negative examples. The self-supervised contrastive learning loss based on neighbor positive samples is calculated as: In formula (11), Remove the positive samples of user u's neighbors from the set U of all users A collection of; for and The cosine similarity between ; for and The cosine similarity between ; τ represents the hyperparameter; Step 3.5: Based on the embedding representation of user u under the target behavior K, supervise the embedding representation of other single behaviors k. The self-supervised contrastive learning loss is: In formula (12), for and The cosine similarity between for and The cosine similarity between is the node representation of multi-behavior perception of user u under target behavior K, is the node representation of user u’s multi-behavior perception under other single behaviors k, is the node representation of multi-behavior perception of user v under other single behaviors k; Step 3.6, under other K-1 different types of single behaviors k, calculate the self-supervised contrastive learning loss of all users as: Step 3.7, repeat step 3.5, and calculate the self-supervised contrastive learning loss for all products as: In formula (14), I is the set of all products, I\{i} is the set of I without product i, is the embedding representation of product i under target behavior K, are the node representations of multi-behavior perception of product i and i' under single behavior k; In step 3.8, the sum of the self-supervised contrastive learning losses of users and products is:

5. The target behavior prediction method based on neighbor context self-supervised contrastive learning according to claim 1 is characterized in that: The specific steps of step S4, calculating the multi-behavior click interest scores of users for the product under different behaviors, optimizing the multi-behavior click interest scores based on the non-sampling strategy, and updating the model weight parameters in conjunction with the self-supervised contrastive learning loss are as follows: Step 4.1, based on the embedding representation dimension d of user u and product i, calculate the click interest score of user u and product i under single behavior k: In formula (16), n is the loop counter of the embedding dimension of users and products; is the value of the nth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (n) is the value of the nth dimension of the embedding representation of a single row k; is the value of the nth dimension of the node representation of multi-behavior perception of product i under single behavior k; Step 4.2: Optimize the click interest score based on the non-sampling strategy. For a user set U consisting of a batch of users b And the product set I consisting of all commodities, the non-sampling loss is defined as: In formula (17), is the set of products that user u interacts with in single behavior k, is the click interest score of user u and product i under single behavior k, Click on the interest score item The positive example weight and negative example weight of m and n are the embedding dimension loop counters of users and products; is the value of the mth dimension of the node representation of the multi-behavior perception of user u under single behavior k; r k (m) is the value of the mth dimension of the embedding representation of a single row k; is the value of the mth dimension of the node representation of multi-behavior perception of product i under single behavior k; Step 4.3, by performing a weighted summation of the non-sampled losses under each single line k, defined as: In formula (18), λ k is the weighting coefficient of non-sampling loss; Step 4.4, combine the unsampled loss and the self-supervised contrastive learning loss through the hyperparameter μ, defined as:

6. The target behavior prediction method based on neighbor context self-supervised contrastive learning according to claim 1 is characterized in that: The step S5, based on the learned user and product node representations, calculates the user's interest score for the new product with respect to the target behavior, sorts the products according to the scores of the target behavior, and generates a product recommendation list in the following specific steps: Embedding r based on target behavior K K , Node representation of user's multi-behavior perception under target behavior K And the node representation of multi-behavior perception of new product i' under target behavior K Calculate the click interest score of user u and new product i' in target behavior K as: In formula (20), is the value of the nth dimension of the node representation of the multi-behavior perception of user u under the target behavior K; r K (n) is the value of the nth dimension of the embedding representation of the target behavior K; is the value of the nth dimension of the node representation of multi-behavior perception of the new product i' under the target behavior K; For multiple new products, calculate the predicted scores of users' interest in different products, and recommend the top t products with higher predicted scores to users.