Hypergraph projection enhanced collaborative filtering

Collaborative filtering (CF) plays a vital role in recommendation scenarios, which models user-item interactions and learns user/item representations to capture correlative patterns in interaction data. Recently, the GNN-based CF models have achieved state-of-the-art performance among various CF methods. Furthermore, some research works try to perform information enhancement of typical GNN-based CF models by extracting global structural semantics with hypergraph structure. However, there are two critical problems with existing methods: (i) Ignorance of local–global correlations during information enhancement and (ii) Insufficient utilization of layer-wise relationships in the residual GNN structure. To address these problems, our proposed Hypergraph Projection Enhanced Collaborative Filtering (HPCF) sufficiently models local–global correlations and layer-wise relationships via representation projection mechanism to adjust the information transfer across them. Thus, our HPCF model can effectively alleviate the negative transfer phenomenon of local–global information enhancement and layer-wise information transfer. Comprehensive experiments on three public datasets demonstrate the state-of-the-art effectiveness of our proposed HPCF as compared with several baselines. The source code and datasets are available at https://github.com/MC-CV/HPCF.