Motif-based Graph Representation Learning for Recommender Systems

Abstract

While graph-based collaborative filtering makes use of the rich user–item relational structure in recommender systems, standard graph neural networks focus mainly on pairwise dependencies, which limits their ability to capture higher-order information that strongly affects recommendation accuracies. This thesis studies how to make graph-based collaborative filtering more robust and powerful by injecting network motifs into graph neural networks. We formalise common bipartite motifs with up to four nodes and demonstrate their integration into representation learning. Building on this, we propose four methods. First, MGSR introduces a motif-level attention mechanism that aggregates motif-induced neighbours in a single layer, which captures higher-order context and outperforming strong baselines on ProgrammableWeb dataset. Second, to scale motif usage beyond small graphs, we design fast generation algorithms for motif adjacency matrices and a lightweight MotifGCN that integrates these matrices into propagation with no extra parameters. Our experiments show that MotifGCN outperform state-of-the-art baselines on four real-world datasets. Through theoretical analysis, we demonstrate that motifs alleviate over-squashing compared with standard layer-wise propagation on original adjacency matrix. Third, we develop MGGCL, a motif-guided contrastive framework that constructs two motif-augmented views so that contrast learning emphasises stable co-interaction signals and alleviates popularity bias introduced by high-degree nodes. Experiment results show consistent performance improvement over baselines, especially on skewed datasets. Finally, we address heterophily with MoHeGCL, which treats each user-anchored triad as a supervision unit, assigns soft homophily labels online, and switches between feature alignment and separation per neighbourhood through heterophily-aware propagation. This improves ranking quality while keeping training costs comparable to SimGCL. Together, these components demonstrate that motif-aware modelling offers a principled way to strengthen message passing, design interpretable self-supervision, and adapt to graphs with different homophily level. We also discuss limitations of our research and outline several future directions.