Recommendations for Contrast Learning Based on Stochastic Masking and Feature-Level Enhancement

Learning graph structure-based representations of user-item interaction data has become the core of modern recommender systems, and graph neural networks (GNNs) show great potential for mining high-quality user-item representations. Therefore, GNN-based collaborative filtering (CF) models have been very successful. However, we believe that further improvement in CF model performance is limited owing to data sparsity and interaction noise in CF. Consequently, the GNN-based CF model is highly prone to problems such as overfitting and poor generalization after multiple graph convolution operations with limited training data. To resolve these problems, we propose a new solution,That is, contrast learning based on stochastic masking and feature-level enhancement(MFCL); specifically, we perform a random masking operation on the resulting embeddings after each layer of graph convolution to prevent the model from learning relations containing excessive sampling noise. In recommender systems, comparative learning better resolves the data sparsity problem than CF because it can extract self-supervised signals from raw data. To further enhance the robustness of the model, we adopt a robust strategy based on feature enhancement. This strategy, called feature blending, blends the feature information of the original graph to obtain an enhanced graph, thereby strengthening the representation learning ability of the nodes. We experimentally demonstrate the superiority and effectiveness of the proposed method using two datasets (MovieLens-1M and Gowalla).