[Paper Review] Heterogeneous Graph Collaborative Filtering

Zekun Li, Yujia Zheng+3 more|arXiv (Cornell University)|Nov 13, 2020

Recommender Systems and Techniques31 references4 citations

TL;DR

This paper proposes Heterogeneous Graph Collaborative Filtering (HGCF), a GCN-based framework that models user-item interactions and user-user behavioral similarity in a unified heterogeneous graph to improve recommendation performance. By explicitly encoding user similarity and enhancing graph connectivity, HGCF reduces reliance on high-order connectivity modeling, achieving state-of-the-art performance with fewer layers and mitigating over-smoothing and sparsity issues.

ABSTRACT

Graph-based collaborative filtering (CF) algorithms have gained increasing attention. Existing work in this literature usually models the user-item interactions as a bipartite graph, where users and items are two isolated node sets and edges between them indicate their interactions. Then, the unobserved preference of users can be exploited by modeling high-order connectivity on the bipartite graph. In this work, we propose to model user-item interactions as a heterogeneous graph which consists of not only user-item edges indicating their interaction but also user-user edges indicating their similarity. We develop heterogeneous graph collaborative filtering (HGCF), a GCN-based framework which can explicitly capture both the interaction signal and similarity signal through embedding propagation on the heterogeneous graph. Since the heterogeneous graph is more connected than the bipartite graph, the sparsity issue can be alleviated and the demand for expensive high-order connectivity modeling can be lowered. Extensive experiments conducted on three public benchmarks demonstrate its superiority over the state-of-the-arts. Further analysis verifies the importance of user-user edges in the graph, justifying the rationality and effectiveness of HGCF.

Motivation & Objective

• To address the limitations of existing bipartite graph-based collaborative filtering methods, which fail to explicitly model user behavioral similarity and suffer from sparsity and high computational cost due to deep high-order connectivity modeling.
• To improve recommendation performance by constructing a more connected heterogeneous graph that integrates both user-item interactions and user-user similarity relations.
• To reduce the demand for deep GCN layers by leveraging the enhanced connectivity of the heterogeneous graph, thereby avoiding over-smoothing and improving efficiency.
• To empirically validate the effectiveness of different similarity measures and graph construction strategies for user-user edge weighting in the heterogeneous graph.

Proposed method

• Construct a heterogeneous graph by enriching a standard user-item bipartite graph with user-user edges based on behavioral similarity, where edge weights are computed using Pointwise Mutual Information (PMI).
• Apply Graph Convolutional Networks (GCN) for end-to-end embedding learning on the heterogeneous graph, enabling joint propagation of interaction and similarity signals.
• Use a two-layer GCN architecture in HGCF, which is sufficient to achieve optimal performance due to the increased connectivity of the heterogeneous graph.
• Compute user-user similarity using PMI, which provides better distinguishability than Cosine or Jaccard similarity by allowing both positive and negative values.
• Apply a similarity threshold to control graph sparsity, balancing overfitting and information retention during graph construction.
• Leverage the heterogeneous graph structure to capture indirect user preferences through lower-order connectivity, reducing computational burden.

Experimental results

Research questions

• RQ1 How does explicitly modeling user-user behavioral similarity in a heterogeneous graph affect recommendation performance compared to traditional bipartite graph-based methods?
• RQ2 What is the optimal similarity measure (e.g., PMI, Cosine, Jaccard) for constructing user-user edges in the heterogeneous graph?
• RQ3 How does the choice of similarity threshold affect model performance and graph sparsity?
• RQ4 Can a shallow GCN (e.g., two layers) on a heterogeneous graph outperform deeper GCNs on a bipartite graph in capturing high-order connectivity?
• RQ5 To what extent does the heterogeneous graph alleviate the sparsity problem and reduce over-smoothing in GCN-based collaborative filtering?

Key findings

• HGCF achieves state-of-the-art performance on three public benchmarks—Gowalla, Yelp2018, and Amazon-Book—outperforming existing SOTA methods.
• Using PMI as the similarity measure yields the best performance across all datasets, with a 0.1614 recall and 0.2330 NDCG on Gowalla, outperforming Cosine and Jaccard.
• The optimal similarity threshold for graph construction is dataset-dependent, with peak performance at t=0 for Gowalla and t=0.5 for Yelp2018, indicating a trade-off between sparsity and information retention.
• HGCF achieves optimal performance with only two GCN layers, while NGCF on the bipartite graph requires three or four layers, demonstrating reduced dependency on high-order connectivity.
• Performance of HGCF drops when more than two layers are used, indicating over-smoothing in highly connected graphs, whereas NGCF shows consistent degradation beyond optimal depth.
• The heterogeneous graph construction significantly reduces the need for deep GCN stacks, proving that enhanced connectivity lowers the demand for high-order modeling while maintaining or improving performance.