Social recommendation model method integrating adaptive graph features and comparison optimization

Abstract

The invention relates to a social recommendation method integrating adaptive graph features and comparison optimization, and discloses a novel social recommendation method integrating adaptive graph features and comparison optimization. By fusing information of a user-article interaction graph and a social graph, the method is innovatively designed for a cold start problem and a data sparsity challenge in social recommendation. The core technical features comprise: (1) a graph fusion feature network structure, dynamically calculating feature weights between nodes and neighbors, effectively fusing features of different graphs, and selectively retaining important features; and (2) a self-adaptive contrast learning framework is adopted, and positive and negative sample pairs are constructed, so that the relevance between the target behavior and the auxiliary behavior can be enhanced, and the sparsity of the target behavior can be relieved. The main purpose of the model is to provide personalized recommendation service for the user, especially in cold start and data sparse scenes, the performance is excellent, the recommendation quality is significantly improved, and the method has high practicability and innovativeness.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of social recommendation systems, and in particular to a recommendation method based on graph neural networks and contrastive learning. BACKGROUND

[0002] With the widespread application of social networks, social recommendation systems have become an important tool to improve user experience. Existing social recommendation systems mainly rely on user-item interaction data and social relationship data, but have the following shortcomings: cold start problem: for new users or new items, due to the lack of sufficient interaction data, traditional recommendation systems are difficult to provide effective recommendations. Data sparsity: the user-item interaction matrix is usually highly sparse, making it difficult for the model to capture rich contextual information and affecting the quality of recommendations. Inadequate use of feature importance: existing methods do not fully consider the importance of different features when integrating social relationships and user behavior, resulting in the loss of some important information. High model complexity: some graph neural network models have high computational complexity when dealing with large-scale data, long training time, and are prone to overfitting. For example, LightGCN simplifies the design of GCN, but still has shortcomings in dealing with cold start problems; SocialLGN introduces social relationships to enhance user representation, but does not effectively address data sparsity and feature importance issues. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application provides a social recommendation system model integrating adaptive graph features and contrastive optimization. The technical solution adopted by the present application is: a dual-graph propagation architecture: through the propagation of the user-item interaction graph and the social graph, the behavior patterns of users and social relationships are captured. An adaptive graph feature fusion network (AGFFN): dynamically adjusts feature weights, enhances feature selectivity and importance, and fuses user representations from the two graphs. An adaptive deep representation contrastive optimization framework (ADRCO): optimizes embedding representations through contrastive learning, improving the discrimination ability and generalization performance of the model. A prediction module: generates recommendation results based on the optimized embedding representations, ensuring that the model can correctly perform high-score user-item interactions. Compared with the prior art, the beneficial effects of the present application are: effectively alleviating the cold start problem: by utilizing social network information to infer the interests of new users, personalized recommendation services are provided for cold start users. Improving recommendation quality: through adaptive feature fusion and contrastive optimization, the accuracy and diversity of recommendations are significantly improved. Enhancing model generalization: through contrastive learning to optimize embedding representations, improving the adaptability of the model on different datasets. Reducing model complexity: simplifying the design of graph neural networks improves the training efficiency of the model. The model of the present application has been experimentally verified on Amazon, Ciao and LastFM datasets, and the results show that the model performs well in cold start and data sparse scenarios, significantly improving the quality of recommendations. BRIEF DESCRIPTION OF DRAWINGS

[0004] In order to make the present application or prior art clearer, the attached drawings needed in the description of the embodiments or prior art will be briefly described in the following: Figure 1 The overall architecture of the social recommendation model proposed in this paper is shown, which integrates adaptive graph features and contrastive optimization. The model input is the user-item interaction graph and the social relationship graph. First, the user and item are initialized to low-dimensional vector representation through the embedding layer. Then, the user embedding is propagated on the social graph and the user-item interaction graph through multi-layer adaptive graph convolutional propagation (AGFFN), and the cross-modal attention mechanism is used to dynamically integrate the information of the two graphs. The fused user / item embedding is then passed through adaptive deep representation contrastive optimization (ADRCO) to further alleviate the data sparsity by constructing positive and negative sample pairs. Finally, the inner product is used to obtain the predicted score and generate Top-K recommendations. Figure 2 The specific calculation steps of the "graph fusion propagation process" in AGFFN are further refined: (1) Input the preprocessed embedding vector (Input Embedding). (2) Map the embedding to a new vector of the same dimension through linear transformation (Linear Transform, parameters are trainable matrix W_g and bias b_g). (3) Use the Sigmoid activation function to generate gate weights (Gate Weights) with a value range of [0,1], which are used to determine the importance of each feature. (4) Perform element-wise multiplication on the gate signal and the original embedding to adaptively enhance or suppress different features while preserving the original information, and output adj_emb. (5) Perform the first L2 normalization on adj_emb to obtain the standardized embedding std_emb, which prevents gradient explosion and stabilizes training. (6) Calculate the attention weight using std_emb through focus identification (Focus Identification) to further generate the weighted embedding. (7) Perform L2 normalization again to obtain the final fused representation, completing the information fusion of the social graph and the user-item interaction graph. DETAILED DESCRIPTION

[0005] The social recommendation system model of the application comprises the following modules: a user-item interaction graph propagation module: propagates the representation of items through a multi-layer graph convolutional network (GCN) to capture the behavior patterns of users. The algorithm receives an initial feature matrix H of users and items, where N is the number of users or items, and F is the dimension of the feature space. Two adjacency matrices represent the user-item interaction graph and the social relationship graph between users, respectively. An embedding list 'embs' is initialized to store the embedding results of each layer. Each layer is processed in a loop, and for each layer, the following operations are performed: combine the user and item embeddings into a unified embedding matrix. This step allows the model to consider both user and item information simultaneously, enabling more effective information propagation. Separate the user and item embeddings. This helps to apply specific operations to different types of nodes in subsequent steps. Update the user embeddings using AGFFN based on the social relationship S between users. AGFFN can capture complex patterns in social networks, improving the accuracy of user representation. Update the user and item embeddings using the user-item interaction adjacency matrix A. This process simulates the interaction between users and items, further enriching the representation of nodes. Separate the updated user and item embeddings again to facilitate the fusion operation in the next step. Then, the embeddings obtained from social network propagation and user-item interaction propagation are fused to generate updated user embeddings. This step ensures that the user representation contains not only social information but also interaction information, enhancing the expressiveness of the model. Update the embedding matrix to prepare for the propagation of the next layer, and add the embedding results of the current layer to the embedding list 'embs'. This allows the accumulation of information from each layer of propagation, ultimately forming a more comprehensive node representation. After all layers of propagation are complete, stack and average the embedding results of all layers to obtain the final user and item embeddings. This step helps to smooth the fluctuations caused by each layer of propagation, stabilizing the final node representation. Finally, separate the user and item embeddings as the output of the algorithm, which can be used for downstream tasks such as the prediction stage of the recommendation system. During the propagation process of each layer, AGFFN layers are called to update the user embeddings, taking into account the social relationships between users. This allows social information to be passed down layer by layer and combined with user-item interaction information, enabling a more comprehensive capture of user behavior patterns and social influences. After updating the social information, it is fused with the user-item interaction information. This multi-layer propagation and fusion approach ensures that social information and interaction information complement each other at different levels, improving the model's understanding of complex user behavior. After multiple layers of propagation and fusion, the updated user and item embeddings are output. The user propagation layer design provides a more comprehensive and accurate representation of users and items by integrating social network information and social interaction information. This deep learning framework not only effectively captures user behavior patterns but also improves the accuracy and personalization of recommendations by utilizing social relationships.Adaptive graph feature fusion network (AGFFN): including feature weight calculation unit and feature fusion unit. The feature weight calculation unit dynamically calculates the feature weight between the node and its neighbor, and identifies the important connection; the feature fusion unit performs weighted fusion on the features of different graphs, and retains important features. Adaptive deep representation contrast optimization framework (ADRCO): including positive and negative sample pair construction unit, similarity calculation unit and contrast loss calculation unit. The positive and negative sample pair construction unit constructs the positive and negative sample pair, and strengthens the relationship between the target behavior and the auxiliary behavior; the similarity calculation unit calculates the similarity between the sample pairs; the contrast loss calculation unit optimizes the embedding representation of the model by shortening the distance of the positive sample pair and lengthening the distance of the negative sample pair. Prediction module: generate ranking score by calculating the inner product of user and item embedding, and combine Bayesian personalized ranking (BPR) loss for optimization to ensure that the model can correctly score high on user-item interaction. In specific implementation, the model is realized by the following steps: initialize the embedding vector of user and item. Perform multi-layer propagation in the user-item interaction graph and social graph respectively to obtain the comprehensive representation of the user. Dynamically adjust the feature weight through AGFFN to fuse the user representation from the two graphs. Optimize the embedding representation through ADRCO to improve the discrimination ability and generalization performance of the model. Generate the recommendation result through the prediction module, and optimize it through the BPR loss. The model of the present application is verified by experiments on Amazon, Ciao and LastFM data sets, and the results show that the model performs well in cold start and data sparse scenarios, and significantly improves the recommendation quality.

Claims

1. A social recommendation method integrating adaptive graph features and contrastive optimization, characterized in that: S1, User-Item Interaction Graph Propagation Module, which is used to propagate the representation of items to capture user behavior patterns; S2, Social Graph Propagation Module, which is used to propagate the user's social relationships to enhance the user's representation; S3, Adaptive Graph Feature Fusion Network (AGFFN), which is used to dynamically adjust feature weights and fuse user representations from two graphs; S4, Adaptive Deep Representation Contrastive Optimization Framework (ADRCO), which is used to optimize the embedding representation through contrastive learning, thereby improving the model's discriminative ability and generalization performance; S5, Prediction Module: The prediction module is used to generate recommendation results based on the optimized embedding representation.

2. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: In S1 and S2, the user-item interaction graph propagation module and the social graph propagation module propagate information through a multi-layer graph convolutional network (GCN). After each layer of propagation, a nonlinear activation function is introduced to introduce nonlinearity, thereby enhancing the expressive power of the model.

3. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: In S3, the Adaptive Graph Feature Fusion Network (AGFFN) includes a feature weight calculation unit and a feature fusion unit; The feature weight calculation unit is used to dynamically calculate the feature weights between nodes and their neighbors to identify important connections; the feature fusion unit is used to perform weighted fusion of features from different graphs to retain important features.

4. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: In S4, the Adaptive Deep Representation Contrast Optimization Framework (ADRCO) includes a positive-negative sample pair construction unit, a similarity calculation unit, and a contrast loss calculation unit; the positive-negative sample pair construction unit is used to construct positive-negative sample pairs to strengthen the relationship between the target behavior and the auxiliary behavior; The similarity calculation unit is used to calculate the similarity between sample pairs; the contrast loss calculation unit is used to optimize the embedding representation of the model by bringing positive sample pairs closer and pushing negative sample pairs further apart.

5. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: In S5, the prediction module generates a ranking score by calculating the inner product of user and item embeddings, and optimizes it by combining Bayesian Personalized Ranking (BPR) loss to ensure that the model can correctly give high scores to user-item interactions.

6. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: The model also includes a normalization unit, which performs L2 normalization on the embedding vectors to ensure that embedding vectors of different scales are comparable when calculating similarity, thereby enhancing the model's generalization ability.

7. The social recommendation method integrating adaptive graph features and contrast optimization according to claim 1, characterized in that: The model is suitable for handling the cold start problem. It infers the interests of new users by using social network information and provides personalized recommendation services for cold start users.

Images