Build an explainable next-best-product recommendation system for banking on AWS

This article describes building an explainable next-best-product recommendation system for banking using PyTorch and Amazon SageMaker with a multi-tower neural network architecture and learned attention mechanisms.

• Multi-tower architecture with specialized neural networks for sequences, transactions, demographics, and behavioral data
• GRU-based sequence tower captures temporal product adoption patterns using pack_padded_sequence for variable-length sequences
• Learned attention mechanism fuses tower outputs with per-customer weights, providing built-in explainability for regulators
• AWS Glue ETL normalizes multi-source banking data; SageMaker Processing creates ML features with Dask parallelism
• Feature Importance Module produces per-customer contribution scores from each data tower without post-hoc interpretation
• Deployed on SageMaker with ml.g5.12xlarge GPU instances; supports batch scoring and real-time endpoints
• Achieved strong top-1, top-3, and top-5 accuracy with sequence tower contributing most signal

The multi-tower approach with learned attention provides both high accuracy and regulatory-compliant explainability for banking product recommendations.