NTCF adapts recommendation depth with Ricci curvature, beating graph baselines
Universal propagation depth over-smooths peripheral nodes; NTCF fixes it with curvature-aware scheduling.
A new paper proposes Neural Tree Collaborative Filtering (NTCF), a framework that reinterprets graph collaborative filtering as tree collaborative filtering. Instead of using a uniform propagation depth for all nodes, NTCF assigns a node-specific depth based on a closed-form local-degree-imbalance score that acts as a discrete Ricci-curvature proxy. The authors prove NTCF strictly generalizes NGCF and retains more discriminative information on positively curved peripheral nodes at deep layers than uniform-depth propagation. Experiments on three public datasets show NTCF achieves higher performance than most widely used GCF backbone models and can also replace the backbone in existing self-supervised models for enhanced performance. The paper is accepted by CIKM 2026 Short.
- NTCF assigns node-specific propagation depth via a discrete Ricci-curvature proxy, unlike fixed-depth GCF models
- Theoretically generalizes NGCF, retaining more discriminative info on peripheral nodes and preventing over-smoothing
- Outperforms standard GCF backbones on three public datasets and boosts self-supervised models when used as a backbone
Why It Matters
Curvature-aware depth tuning could improve accuracy in large-scale recommenders with minimal added cost or complexity.