Feature Re-enhanced Meta-Contrastive Learning for Recommendation

Enhancing the performance of recommendation systems through joint modeling of user-item interactions and knowledge graph (KG) information using Graph Neural Networks (GNN) has shown promising results. However, due to the cold-start problem and inherent sparsity in graph data, the effectiveness of extracting informative features from the graph data is often limited. Existing approaches mostly focus on enhancing the original embeddings. It is imperative to develop a comprehensive and effective multi-faceted feature enhancement strategy that assists in augmenting information during the feature transfer process. In this paper, we present a novel framework for recommendation called Feature Re-enhanced Meta Contrastive Learning (FReML). The framework employs a simple self-gating mechanism to handle the features in the interaction view, enhancing the correlation between user and item features. It utilizes a multi-layer perceptron approach for meta-feature extraction on both the interaction view and the knowledge graph view, aiming to enhance the performance of contrastive learning. We evaluate our proposed framework on real-world datasets and demonstrate improved training results compared to state-of-the-art approaches. Further ablation analysis validates the rationality of the key component designs.