NetEase's NDL proactively denoises graph collaborative filtering via representation coherence
_reachsumit · x · 2026-11-11
A new arXiv paper from NetEase researchers proposes Noise-aware Dimension Learning (NDL) for graph collaborative filtering:
- Problem: noisy implicit interactions propagate through user-item graphs; existing denoising relies on training loss, a passive signal that detects noise only after prediction errors and is confounded by training stage, negative sampling, and hard positives.
- Method: Representation Coherence Detection (RCD) measures an interaction's coherence with the user's preference centroid in embedding space as an active noise signal; fused with loss via activity-aware weighting, it drives an interaction-level dimension-adaptive mask, plus a contrastive regularizer for embedding uniformity.
- Results: consistent improvements over existing methods on three public benchmarks.
More from Research
- Ethan Mollick has AI run a Fermi paradox study on public data: 73% odds we're alone in the Milky Way — emollick · 2026-11-11
- DNAForge & MTC Berkeley to host AI × Biology Hackathon in Berkeley on Nov 14 — julia_kiseleva · 2026-11-11
- RoboScript hits 83.4% success on RoboTwin 2.0 with one demo per task, beating trained policies by 20+ points — H-EmbodVis · 2026-11-11
- Alibaba's PhysOPSD uses executable code as privileged context to make video generation physically plausible — AlibabaTongyiLab · 2026-11-11
- mayfer: clean low-rank latents may be the key to continuous online learning — mayfer · 2026-11-11
- PhD student finds lost Mayan city with 6,000+ buildings in old LiDAR dataset — bilawalsidhu · 2026-11-11