Reproduce “Blurring-Sharpening Process Models for Collaborative Filtering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jeongwhan Choi, Seoyoung Hong, Noseong Park, Sung-Bae Cho
Blurring-Sharpening Process Models (BSPM) extends parameter-free graph-based collaborative filtering by pairing conventional interaction smoothing with an ODE-based sharpening stage. Across Gowalla, Yelp2018, and Amazon-book, the method achieved the best reported Recall@20 and NDCG@20 among 43 baselines, while also improving item coverage, novelty, and the recommendation of less popular items.
Key insights
- Moderate sharpening can recover user-specific information that graph diffusion suppresses; excessive sharpening instead removes useful collaborative signal, making the terminal sharpening time important.
- The gains extend beyond ranking accuracy: sharpening increases coverage and novelty and shifts correct recommendations toward lower-degree, less popular items.
- The continuous-time formulation unifies or recovers behavior resembling several existing filters, including LightGCN-like diffusion and GF-CF-style operations, while adding a complementary recovery process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.