Reproduce “Personalised Outfit Recommendation via History-aware Transformers” 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
Myong Chol Jung, Julien Monteil, Philip Schulz, Volodymyr Vaskovych
Amazon Machine Learning introduces the History-aware Transformer (HAT), a model designed to provide personalized outfit recommendations by effectively integrating shopper purchase history. It achieved substantial performance gains, with up to a 19.4% AUC increase on challenging personalized outfit compatibility prediction tasks and a 9.7% accuracy improvement on fill-in-the-blank tasks.
Key insights
- Integrating a shopper's historical outfit purchases via a top-level transformer architecture is crucial for accurately inferring and applying individual fashion tastes.
- Supervised contrastive learning effectively organizes outfit embeddings, making styles from the same shopper more similar, which is fundamental for personalized recommendations.
- The use of "weak negatives" (outfits with a single item replaced) and an adaptive margin loss weighted by item importance yields a refined training signal for distinguishing subtle compatibility variations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.