Part 7/12:

Each product is mapped to a unique token ID, enabling the model to process purchase sequences seamlessly, akin to processing text.

Handling Data and Seasonality

The model was trained on 8-week rolling windows of purchase history, with 52 such windows per customer to account for seasonal behaviors. For example, one training sample might consist of purchase data from January-February, with the target being products bought in March. Importantly, only first-time purchases were included in the target sequences, capturing the model’s ability to recommend novel items.

Implementation Details and Training

• The team trained a compact T5 model on NVIDIA T4 GPUs, carefully tuning parameters to fit resource constraints.