Long-tail recommendation method based on two-channel adaptive graph convolutional network

Abstract

The invention discloses a long-tail recommendation method based on a dual-channel adaptive graph convolutional network, and belongs to the technical field of recommendation systems. In order to solve the problem of recommendation deviation caused by sparse interaction data of long-tail users in an existing recommendation system, the invention provides a long-tail recommendation method of a dual-channel adaptive graph convolutional network. Firstly, an interaction matrix is constructed according to a user-project bipartite graph, a dual-channel representation learning architecture is adopted, feature vectors of head users of high-frequency interaction and tail users of sparse interaction are modeled respectively, and representation overlapping is reduced; secondly, dynamically adjusting the propagation weight of a low-degree node (a tail user or an item) through a degree perception weighted aggregation mechanism, enhancing the expression capability of the low-degree node, and inhibiting the excessive smoothing effect of a high-interaction node; and finally, generating a personalized recommendation score by fusing two-channel prediction results. According to the method, the accuracy and fairness of long-tail user recommendation are remarkably improved, and the method is suitable for personalized recommendation requirements of scenes such as e-commerce and social media.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and in particular to a long-tail recommendation method based on a dual-channel adaptive graph convolutional network, which is suitable for optimizing personalized recommendations for long-tail users and items in scenarios such as e-commerce and social media. Background Art

[0002] With the rapid development of the internet and big data technologies, recommendation systems have been widely used in e-commerce, social media, and content distribution platforms. Recommendation systems generate personalized recommendations by mining user-item interaction data, thereby improving user experience. However, real-world user-item interaction data typically exhibits a long-tail distribution, where a small number of head users and popular items account for the majority of interaction records, while the interaction records of a large number of tail users and niche items are extremely sparse. This long-tail distribution makes it challenging for recommendation systems to provide accurate recommendations for tail users and items. Sparse data often leads to insufficient model representation, reduced recommendation quality, and impacts user trust and the coverage of personalized services.

[0003] In recent years, recommendation methods based on graph convolutional networks (GCNs) have improved recommendation performance by modeling high-order neighbor relationships in user-item interaction graphs. For example, LightGCN (Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the arXiv preprint arXiv:2002.02126.) improves prediction accuracy and computational efficiency by simplifying graph convolution operations, removing feature transformations and nonlinear activations. However, LightGCN has limitations when processing long-tail data: its propagation mechanism uniformly weights all nodes, ignoring the differences in interaction frequencies between head and tail nodes. This results in the dilution of tail user and item features due to sparse interactions, resulting in degraded recommendation performance. Furthermore, LightGCN lacks an adaptive mechanism to balance the contributions of head and tail nodes, limiting its effectiveness in long-tail recommendation scenarios.

[0004] Recently, various solutions have been proposed for the long-tail recommendation problem. For example, GALORE (Sichun Luo, Chen Ma, Yuanzhang Xiao, and Linqi Song. 2023. Improving Long-Tail Item Recommendation with Graph Augmentation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM'23). 1707-1716.) uses graph augmentation technology to increase the connectivity of tail nodes, promote information transfer from head to tail nodes, and adopts adaptive edge dropping and synthetic node generation to alleviate data sparsity. Although GALORE outperforms LightGCN in tail item recommendation, it mainly focuses on graph structure adjustment and does not fully utilize the interactive feature distribution between head and tail nodes, which limits further improvement in the quality of tail user and item recommendations. In addition, other methods such as resampling, regularization, similarity-based random walks, and re-transfer learning, although they have optimized long-tail recommendations to a certain extent, have failed to fully address the core problem of insufficient representation of tail nodes. Especially in graph-based recommendation scenarios, the propagation-dominant effect of head nodes still significantly affects the representation quality of tail nodes.

[0005] This paper proposes a long-tail recommendation method based on a dual-channel adaptive graph convolutional network. Through dual-channel representation learning and degree-aware weighted aggregation mechanism, it optimizes the feature expression of head and tail users and items respectively, overcomes the shortcomings of LightGCN's uniform weighting, and more fully explores the interactive feature distribution than GALORE, improving the recommendation performance and fairness of tail users and items while maintaining computational efficiency. Summary of the Invention

[0006] Purpose of the invention: In response to the problems of low quality, insufficient representation and poor fairness of tail user and project recommendations faced by recommendation systems in background technologies in long-tail scenarios, the present invention proposes a long-tail recommendation method based on a dual-channel adaptive graph convolutional network, aiming to improve the recommendation performance and fairness of tail users and projects while maintaining computational efficiency through dual-channel representation learning and degree-aware weighted aggregation mechanism.

[0007] Technical solution: To achieve the above objectives, the present invention adopts the following technical solution: a long-tail recommendation method based on a dual-channel adaptive graph convolutional network, comprising the following steps:

[0008] Step S1: Construct user-item interaction graph and training dataset;

[0009] Based on user-item interaction data, a user-item interaction graph is constructed, in which nodes include user nodes and item nodes, and edges represent the interaction relationship between users and items; according to the interaction frequency, users are divided into top users (top 20%) and tail users (bottom 80%) according to the Pareto principle, and items are divided into top items (top 20%) and tail items (bottom 80%), and a training dataset is constructed based on the interaction graph and interaction matrix.

[0010] Step S2: Initialize the dual-channel adaptive graph convolutional network model;

[0011] Initialize the dual-channel adaptive graph convolutional network, including the initial embedding vectors of the head channel and the tail channel. The embedding dimension is set to 64 to balance the model's expressiveness and computational efficiency. Define the head channel embedding for user node u and item node i respectively. and tail channel embedding The head channel is used to capture the high-frequency interaction features of head users and items, and the tail channel is used to optimize the sparse interaction features of tail users and items.

[0012] Step S3: perform multi-layer graph convolution and degree-aware weighted aggregation;

[0013] Three layers of standard graph convolution are performed in the head and tail channels respectively to update the node embeddings. Then, weight coefficients are calculated based on the node degrees, and a degree-aware weighted aggregation mechanism is applied to enhance the propagation contribution of tail users and items. Specifically,

[0014] (1) Calculating user nodes and project nodes Degree and in and are the neighbor node sets of users and items respectively;

[0015] (2) Define the degree perception weighting coefficient:

[0016]

[0017] (3) l=0,1,2, the update formula is as follows:

[0018]

[0019] (4) After the third layer, degree-aware weighted aggregation is applied to calculate the final embedding:

[0020]

[0021] Step S4: Fusion of dual-channel prediction results;

[0022] The final recommendation score is calculated by weighted fusion of the interactive prediction scores of the head channel and the tail channel.

[0023]

[0024] Where W head , w tail is the learnable weight coefficient, λ tail is the tail channel amplification factor, which is used to enhance the contribution of tail user and item features and is optimized through hyperparameter search.

[0025] Step S5: model optimization and recommendation output;

[0026] The Bayesian Personalized Ranking (BPR) loss function is used to optimize the model, combined with the embedding regularization loss. The specific loss function is:

[0027]

[0028] Where σ is the sigmoid function, λ is the regularization coefficient, and (u, i, j) indicates that user u prefers the training triplet of positive sample item i rather than negative sample item j. Through iterative optimization until the model converges, the recommendation result for the user is output.

[0029] Furthermore, in step S2, the embedding dimension is set to 64 to balance the model's expressiveness and computational efficiency.

[0030] Furthermore, in step S3, the number of graph convolution layers is preferably 3 to strike a balance between capturing high-order neighbor relationships and avoiding over-smoothing.

[0031] Furthermore, in step S4, the tail channel amplification coefficient λ is determined by hyperparameter search to optimize the recommendation performance of tail users and items.

[0032] Beneficial effects

[0033] The present invention introduces dual-channel representation learning to model the high-frequency interaction features of head users and projects and the sparse interaction features of tail users and projects respectively, effectively avoiding the dilution problem of tail user and project features being dominated by head interactions, thereby improving the recommendation quality and personalized service coverage of tail users and projects.

[0034] The degree-aware weighted aggregation mechanism proposed in this invention enhances the representation ability of tail users and items by dynamically adjusting the node propagation weights, suppresses the over-smoothing effect of head users and items, and improves the fairness of recommendations.

[0035] While maintaining computational efficiency, the present invention is compatible with the simplicity of existing graph convolutional networks and is suitable for large-scale sparse interaction scenarios. It has strong versatility and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 :The algorithm framework diagram of the long-tail recommendation method based on the dual-channel adaptive graph convolutional network, which includes the following elements:

[0037] 1. Input layer: Displays the user-item interaction graph and interaction matrix, and labels the node classification (top user / tail user, top item / tail item).

[0038] 2. Dual-channel structure: It is divided into a head channel and a tail channel, which respectively display the initial embedding of users and items and the three-layer graph convolution process.

[0039] 3. Degree-aware weighted aggregation: Marks the user and project degree calculation and weighting coefficient formula, highlighting the weight increase of tail users and projects.

[0040] 4. Fusion layer: Displays the weighted fusion of user and item embeddings from the head and tail channels, and outputs the recommendation score.

[0041] 5. Loss function: Annotate the optimization process of Bayesian personalized ranking (BPR) loss and regularization loss.

[0042] Figure 2 : Flowchart of method execution, including the following modules:

[0043] 1. Data preparation: Build a user-item interaction graph, divide the top / tail users and items, and generate a training dataset.

[0044] 2. Model initialization: Initialize the two-channel adaptive graph convolutional network and set the user and item embeddings of the head channel and tail channel.

[0045] 3. Graph Convolution and Weighted Aggregation: Perform three layers of graph convolution and apply degree-aware weighted aggregation to generate the final embeddings of users and items.

[0046] 4. Prediction fusion: Fusion of the embeddings of the head channel and the tail channel to calculate the recommendation score.

[0047] 5. Model optimization and output: Use BPR loss and regularization loss to optimize the model and output personalized recommendation results. DETAILED DESCRIPTION

[0048] The present invention provides a long-tail recommendation method based on a dual-channel adaptive graph convolutional network. The specific algorithm framework is as follows: Figure 1 As shown, the process is as Figure 2 As shown, the following steps are included:

[0049] Step S1: Construct a user-item interaction graph and training dataset.

[0050] Based on historical user-item interaction data from e-commerce or social media platforms, we construct a user-item interaction graph, with users and items as nodes and interaction records as edges. Based on interaction frequency, we apply the Pareto principle to categorize users into top users (the top 20%) and bottom users (the bottom 80%), and items into top items (the top 20%) and bottom items (the bottom 80%). We then generate a training dataset based on the interaction graph and interaction matrix.

[0051] Step S2: Initialize the dual-channel adaptive graph convolutional network model.

[0052] Initialize the dual-channel adaptive graph convolutional network model and set the initial embedding vectors of the head channel and the tail channel The embedding dimension is set to 64 to balance the model's expressiveness and computational efficiency. The head channel captures the high-frequency interaction features of the head users and items, while the tail channel optimizes the sparse interaction features of the tail users and items.

[0053] Step S3: Perform multi-layer graph convolution and degree-aware weighted aggregation.

[0054] Perform 3 layers of standard graph convolution in the head and tail channels respectively to update the node embedding. Calculate the degree d of user and item nodes u d i , and generates weighting coefficients based on the formula.

[0055]

[0056] Step S4: Fusing the dual-channel prediction results.

[0057] Through the weighted fusion formula Calculate the recommendation score, where W head 、w tail Through model training, λ tall Optimization via hyperparameter search.

[0058] Step S5: Model optimization and recommendation output.

[0059] The BPR loss and regularization loss are used to optimize the model. After iterative training until convergence, a personalized recommendation list for the user is output.

Claims

1. A long-tail recommendation method based on a dual-channel adaptive graph convolutional network, characterized in that: The following steps are involved: Step S1: Obtain user-item interaction data, construct a user-item interaction graph, and divide users and items into top users and tail users, top items and tail items based on the interaction frequency according to the Pareto principle to generate a training dataset; Step S2: Initialize the model parameters of the dual-channel adaptive graph convolutional network, including the initial user embedding and item embedding of the head channel and the tail channel; Step S3: Perform three-layer graph convolution operations in the head channel and the tail channel respectively to independently model the interaction features of the head user and item and the sparse interaction features of the tail user and item; Step S4: Apply degree-aware weighted aggregation mechanism based on node degree to calculate the final user embedding and item embedding of the head channel and tail channel; Step S5: Fuse the embeddings of the head channel and the tail channel to generate a user-item recommendation prediction score and output the recommendation result.

2. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The method for constructing the user-item interaction graph and dividing users and items in step S1 includes: Generate a user-item interaction graph based on user-item interaction data, where users and items are respectively used as nodes and interactions as edges; Based on the interaction frequency, users are divided into top users (top 20%) and bottom users (bottom 80%) according to the Pareto principle, and projects are divided into top projects (top 20%) and bottom projects (bottom 80%). Generate training dataset based on user-item interaction graph and interaction matrix.

3. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The method for initializing the model parameters of the dual-channel adaptive graph convolutional network in step S2 includes: and item i∈I respectively initialize the head channel embedding and tail channel embedding The embedding dimension d is 64 to balance the model’s expressiveness and computational efficiency; Head channel embedding is used to capture the high-frequency interaction features of head users and items, and tail channel embedding is used to capture the sparse interaction features of tail users and items.

4. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The specific steps of the multi-layer graph convolution operation in step S3 include: In the head channel and tail channel, three layers of standard graph convolution operations are performed on the user node and item node respectively, with the number of layers l = 0, 1, and 2. The update formula is as follows: in, are the user embeddings of the head channel and tail channel of the l+1 layer respectively, are the item embeddings of the head channel and tail channel of the l+1 layer respectively, is the set of neighbor items for user u, is the set of neighbor users of project i.

5. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The specific steps of the degree-aware weighted aggregation mechanism in step S4 include: Step S4.1, calculate the degree of user u and item i, respectively and Step S4.2: Calculate the weighting coefficient based on the degree: Among them, w deg (u), w deg (i) Degree weighting coefficients for users and items respectively; Step S4.3, after the third layer of graph convolution, apply the degree weighting coefficient to calculate the final embedding:

6. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The method for fusing the head channel and the tail channel embedding to generate the recommendation prediction score in step S5 includes: By weighted fusion of the embeddings of the head channel and the tail channel, the prediction score of user u for item i is calculated The formula is as follows: Among them, w head 、w tail is the learnable weight, is the tail channel amplification coefficient, λ tail is the tail channel amplification factor, which is used to enhance the contribution of tail user and item features and is determined through hyperparameter search.

7. The long-tail recommendation method based on a dual-channel adaptive graph convolutional network according to claim 1, characterized in that: The model optimization of the method adopts the Bayesian personalized ranking loss function, which is formulated as follows: Where (u, i, j) represents the training triplet in which user u prefers positive sample item i over negative sample item j, and σ is the sigmoid function; At the same time, embedding regularization loss is introduced: The total loss is: Where λ is the regularization coefficient.

Images