Part 8/12:
- During training, the model learned to generate the probability distribution over all products given an input sequence, enabling ranking by likelihood.
Training stability was achieved after approximately 200 epochs, with cross-entropy loss stabilizing, confirming effective learning.
Results and Performance Gains
When evaluated against the ALS baseline, the Basket Transformer achieved:
Approximately 15% improvement in precision at the 10th and 20th recommendation slots.
Better rank ordering of suggested products, evidenced by higher decimal gains metrics.
This demonstrated the model's superior ability to not only predict what a customer might buy but also to prioritize those recommendations effectively.