Context-aware Graph Collaborative Recommendation Without Feature Entanglement

Abstract

Inheriting from the basic idea of latent factor models like matrix factorization, current collaborative filtering models focus on learning better latent representations of users and items, by leveraging the expressive power of deep neural networks. However, the settings where rich context information is available pose difficulties for the existing neural network based paradigms, since they usually entangle the features extracted from both IDs and other additional data (e.g., contexts), which inevitably destroy the original semantics of the embeddings. In this work, we propose a context-aware collaborative recommendation framework called CGCR to integrate contextual information into the graph-based embedding process. Our model converts the bipartite graph to a homogeneous one by placing the users and items in the identical feature space. As our method is free of feature crosses, it can preserve the semantic independence on the embedding dimensions and thus improves the interpretability of neural collaborative filtering. We use generalized matrix factorization as the matching function so that the model can be trained in an efficient non-sampling manner. We further give two examples of CGCR: LGC with linear graph convolutional networks and LGC+ with attention mechanism. Extensive experiments on five real-world public datasets indicate that the proposed CGCR models significantly outperform the state-of-the-art methods on the Top-K recommendation task.