Reproduce “Turbo-CF: Matrix Decomposition-Free Graph Filtering for Fast Recommendation” 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
Jin-Duk Park, Yong-Min Shin, Won-Yong Shin
Turbo-CF replaces the costly spectral decomposition used by graph-filtering recommenders with low-order polynomial filters over an item–item graph. Across Gowalla, Yelp, and Amazon-book, it preserves accuracy close to GF-CF and LightGCN while reducing Gowalla runtime to 0.32–5.79 seconds, compared with hours for trained GCN models and minutes for GF-CF.
Key insights
- Most of the practical speedup comes from avoiding matrix decomposition, but higher-order filters can still become memory-intensive because they require additional item–item matrix operations.
- More filtering complexity does not guarantee better recommendations: the best order varies by dataset, with third-order, second-order, and linear variants preferred on Gowalla, Yelp, and Amazon-book, respectively.
- Graph construction choices matter alongside the filter itself; on Gowalla, asymmetric normalization with alpha = 0.7 and edge adjustment with s = 0.6 outperform the symmetric or unadjusted settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.