Reproduce “Recommender System as Slow and Fast Thinkers” 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
Zichen Yuan, Xiaoxuan Dong, Linkun Dai, Jinwei Yang, Jining Luan, Dexu Yu, Chunxiao Li, Joemon M. Jose and 3 more
Researchers from City University of Hong Kong and collaborating institutions introduce DS-Frame, an adaptive inference framework for sequential recommender systems that dynamically allocates computational resources using 'fast' and 'slow' processing paths based on prediction difficulty. It consistently improves recommendation performance, particularly for users with complex interaction histories, while maintaining efficiency by selectively invoking deeper inference.
Key insights
- Empirical analysis reveals that static sequential recommenders exhibit significant performance disparities across diverse user groups, with notable degradation for challenging environments.
- An adaptive computation strategy, inspired by Dual Process Theory, can effectively improve recommendation quality by selectively allocating more computational resources to complex prediction instances.
- A learned routing mechanism can successfully identify and direct challenging user sequences to a more expressive 'slow' reasoning path, thereby optimizing the accuracy-efficiency trade-off under a controllable budget.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.