Collaborative Filtering Algorithm Based on Contrastive Learning and Filtering Components

The recommendation system based on graph neural networks has attracted significant attention due to its powerful learning capability, especially in systems based on social network graphs. Combining social network graphs with user-item graphs can capture dynamic preference features, achieving more accurate recommendations. However, user behavioral data often contains noise and is sparse, which may result in suboptimal model performance. To address this issue, a Collaborative Filtering Model based on Contrastive Learning and Filtering Components (FMPRec) is proposed. Drawing inspiration from filtering algorithms in signal processing, this model reduces noise in the frequency domain. By combining learnable filter components with a fully MLP architecture, it learns the latent embeddings of users and items, and predicts ratings through the fully MLP structure, with lower time complexity. Additionally, leveraging graph structure and semantic space introduces neighbor information of users and items, fully exploiting the potential of contrastive learning. Experimental results demonstrate that the proposed approach outperforms traditional methods in terms of MAE/RMSE metrics on two real datasets, validating its feasibility and effectiveness in collaborative filtering research.