Reproduce “Balancing Trial and Reorder: A Hybrid Sequential Transformer-GBDT Ranker for On-Demand Delivery” 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

Marcel Kurovski, Attila Nagy, Steffen Klempau, Aleksandr Fedintsev

The UVR system ranks eligible delivery stores by combining a bidirectional transformer’s representation of a customer’s purchase history with store, user–store, and delivery-context features in separate restaurant and retail GBDT rankers. Per-session weights tune the balance between first-time and repeat purchases. In randomized tests, V1 increased Merchant Trial Rate 5.5% and Global CVR 0.16%; later CVR changes were not significant.

Key insights

  • The discovery objective can be adjusted operationally through the ratio of new-store to reorder training weights; increasing this emphasis improves trial metrics but tends to weaken reorder performance offline.
  • The transformer captures useful cross-domain purchase-history signals, while the GBDT adds complementary operational and user–store information and outperforms the transformer alone across all reported comparisons.
  • A unified architecture can simplify a fragmented production stack, but transductive store and location embeddings make daily retraining important and leave newly introduced stores reliant on content and contextual features.

Summary by alphaXiv.

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