Next-Basket Prediction using Bidirectional Transformers

Supervisor: Dr. Yucel Saygin

Next-basket prediction asks: given a customer’s purchase history, what items will appear in their next basket? Sequential models tend to over-weight recency; bidirectional attention captures longer-range complementarity (e.g. “bought a printer three months ago, likely to need ink now”).

We applied a BERT-style bidirectional Transformer with basket-level masking and augmented training data to address sparsity in the long tail of SKUs. The model improved hit-rate@K on held-out baskets over sequential-only baselines.