Reproduce “Neural Tree Collaborative Filtering: Rethinking Graph Collaborative Filtering as Tree Collaborative Filtering with Curvature-Aware Propagation Depth” on OpenResearch

About this paper

Jinfeng Xu, Zheyu Chen, Ziyue Peng, Shuo Yang, Jinze Li, Wenhao Yuan, Jian Chen, Edith C. H. Ngai

Neural Tree Collaborative Filtering (NTCF) addresses the limitation of uniform message propagation depth in Graph Collaborative Filtering (GCF) models by dynamically assigning a node-specific propagation depth based on local connectivity patterns. This adaptive approach consistently outperformed six GCN-based baselines and improved four self-supervised recommendation models across three datasets.

Key insights

• Adopting a node-adaptive propagation depth based on local graph geometry allows for a more flexible and effective aggregation of collaborative signals in GCF.
• The local degree-imbalance score provides an efficient and meaningful metric to predict the optimal propagation depth for individual nodes.
• NTCF strictly generalizes existing GCF methods like NGCF and theoretically mitigates the over-smoothing phenomenon by preserving discriminative information for peripheral nodes at deeper layers.

Summary by alphaXiv.