Reproduce “Sliding Window Training -- Utilizing Historical Recommender Systems Data for Foundation Models” 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
Swanand Joshi, Yesu Feng, Ko-Jen Hsiao, Zhe Zhang, Sudarshan Lamkhede
Netflix researchers developed a sliding window training method for recommender system foundation models, allowing them to learn from extended user interaction histories (up to 1000 items) without increasing inference-time input sequence length. This approach consistently improved next-item prediction accuracy and item representation quality across various metrics compared to fixed-window training.
Key insights
- Exposing recommender models to a broader range of historical user interactions during training significantly enhances their learning capabilities, even if only a limited recent window is used for inference.
- A mixed training strategy that combines recent interaction focus with historical sliding windows provides a balanced approach that outperforms purely historical or purely recent training.
- Improved exposure to diverse item interactions, including less frequent ones, leads to better-learned item representations, beneficial for tasks like item cold-start.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.