Abstract
Deep learning on graphs, specifically graph convolutional networks (GCNs), has exhibited exceptional efficacy in the domain of recommender systems. Most GCNs have a message-passing architecture that enables nodes to aggregate information from neighbours iteratively through multiple layers. This enables GCNs to learn from higher-order information, but the model does not allow for direct captions of the local structural patterns. Our rationale is to investigate the effectiveness of capturing such local patterns for graph-based collaborative filtering to enhance model’s learning ability per layer. This technique combines lower-order and higher-order interactions during layer-wise propagation. In this paper, we propose MotifGCN to aggregate both lower-order and higher-order information in each graph convolution layer. Specifically, we develop dedicated algorithms of generating motif adjacency matrices. The matrices are then used for motif-enhanced neighbourhood aggregation in each layer. As this paper focuses on recommender systems, MotifGCN is built on the basis of bipartite graphs. Our experiments on four real-world datasets show that MotifGCN has a superior performance compared to various state-of-the-art methods.
| Original language | English |
|---|---|
| Pages (from-to) | 9413-9432 |
| Number of pages | 20 |
| Journal | Neural Computing and Applications |
| Volume | 37 |
| Issue number | 16 |
| DOIs | |
| Publication status | Published - Jun 2025 |
Bibliographical note
Copyright the Author(s) 2025. Version archived for private and non-commercial use with the permission of the author/s and according to publisher conditions. For further rights please contact the publisher.Keywords
- Recommender systems
- Collaborative filtering
- Graph convolutional networks
- Network motifs
Fingerprint
Dive into the research topics of 'Improving graph collaborative filtering with network motifs'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver