Reproduce “Language-Model Prior Overcomes Cold-Start Items” 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

Shiyu Wang, Hao Ding, Yupeng Gu, Sergul Aydore, Kousha Kalantari, Branislav Kveton

Researchers at AWS AI Labs developed a Bayesian framework that integrates Language Model-derived item similarities as a graph regularizer within existing recommender systems. This approach robustly enhances recommendation performance, particularly for cold-start items, showing an average of 40% higher NDCG@10 on MovieLens and Amazon cold-start items with localized similarity estimation compared to global methods.

Key insights

  • An LM-derived Bayesian prior, applied as a graph regularizer, effectively guides the learning of item embeddings, leading to robust performance for both established and cold-start items.
  • Capturing localized semantic relationships through item-specific covariance estimation in the LM embedding space yields superior performance over a single global variance parameter, demonstrating the importance of fine-grained similarity.
  • The approach is a generic, "plug-and-play" framework that can enhance various recommender architectures (e.g., BPRMF, SASRec) by leveraging unstructured textual data without requiring major system overhauls.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.