Reproduce “MvFS: Multi-view Feature Selection for Recommender System” 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
Youngjune Lee, Yeongjong Jeong, Keunchan Park, SeongKu Kang
Researchers from NAVER Corporation and the University of Illinois at Urbana-Champaign introduced Multi-view Feature Selection (MvFS), an adaptive method designed to mitigate feature selection bias and inter-feature dependencies in recommender systems. MvFS consistently improved AUC by up to 0.0061 and reduced Logloss across four different recommender models and two real-world datasets, outperforming prior adaptive selection techniques.
Key insights
- A multi-view architecture, where distinct sub-networks specialize in different feature patterns, effectively mitigates bias towards major features and enhances the consideration of minor, yet informative, features.
- Modeling feature importance scores independently for each field, without global reweighting, is crucial for preventing spurious dependencies and accurately reflecting individual feature contributions.
- A gradual transition from soft to hard feature selection during training, using a smoothed step function, effectively balances exploration and exploitation, leading to more robust feature selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.