Neural Graph Collaborative Filtering

Neural Graph Collaborative Filtering (NGCF) targets a drawback in recommender systems where learning user and item embeddings, from early matrix factorization to deep learning methods, typically maps from pre-existing features like IDs and attributes and does not encode the collaborative signal latent in user-item interactions. The authors argue such embeddings may be insufficient to capture the collaborative filtering effect, so they integrate the user-item interactions, specifically the bipartite graph structure, into the embedding process by propagating embeddings over the graph. This yields expressive modeling of high-order connectivity, explicitly injecting the collaborative signal.

Extensive experiments on three public benchmarks demonstrate significant improvements over several state-of-the-art models, including HOP-Rec and Collaborative Memory Network. Further analysis verifies the importance of embedding propagation for learning better user and item representations, justifying the approach's rationale and effectiveness. The work, with released code, helped establish graph-based propagation as a strong paradigm for collaborative filtering.