Reproduce “Adaptive Quality-Diversity Trade-offs for Large-Scale Batch 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
Clémence Réda, Tomas Rigaux, Hiba Bederina, Koh Takeuchi, Hisashi Kashima, Jill-Jênn Vie
Researchers from École Normale Supérieure PSL, Kyoto University, and Inria developed a scalable and adaptive framework for large-scale batch recommendation that optimizes for both relevance and diversity. Their B-DivRec algorithm, built on an extended Determinantal Point Process (DPP) formulation, integrates user history for global diversity and uses an online learning approach to dynamically adjust the quality-diversity trade-off for individual users, demonstrating superior performance and efficiency across diverse datasets.
Key insights
- Determinantal Point Processes can be effectively extended and made computationally scalable for large-scale batch recommendations by carefully disentangling quality and diversity components and using approximations like Nyström and truncated SVD.
- Integrating user history through a 'fuzzy denuding' mechanism within a DPP framework allows for nuanced global diversity without significantly compromising relevance.
- Online, adaptive tuning of the quality-diversity trade-off provides personalized recommendations that respond to changing user preferences and improves overall performance, even when feedback is noisy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.