Build an explainable next-best-product recommendation system for banking on AWS
AWS Ayush Singh Chauhan
AWS published a blueprint for banks to predict a customer's next likely product using deep learning on SageMaker. It bakes in explainability, not bolted-on SHAP tricks, which matters a lot for regulators.
Based on reporting by AWS, Ayush Singh Chauhan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Banks sit on mountains of customer data—transaction logs, product histories, demographics, behavioral tags—but turning that into a useful "here's what to offer this person next" system has always been messier than it sounds. AWS's latest architectural writeup tackles this with what they call a Next-Best-Product recommender, built on SageMaker AI and PyTorch, and the interesting part isn't the accuracy claims. It's the explainability, which is baked into the model's forward pass rather than bolted on afterward with SHAP or LIME.
The design splits customer data into four separate neural towers, each handling one data type on its own terms. A sequence tower runs a two-layer GRU over a customer's product adoption history—what they bought and in what order—fused with their active product count. Transaction, demographic, and behavioral data each get their own two-layer MLP. AWS's reasoning is straightforward: cramming ordered ID sequences and numerical aggregations through identical layers wastes capacity, since the data shapes are fundamentally different. They picked GRU over LSTM specifically because it has a third fewer parameters and trains faster while performing about the same on sequences of 20 items or fewer, which is roughly what a customer's real product history looks like.
The fusion step is where the explainability actually lives. Instead of just concatenating the four tower outputs, a learned attention mechanism weighs each tower per customer, and those weights sum to something a relationship manager can actually read. A newer customer with thin transaction history but solid demographic data ends up leaning on the customer tower; someone with years of activity leans on the transaction tower instead. AWS adds a Feature Importance Module on top that spits out per-customer scores summing to 1.0 as part of every prediction, no post-hoc interpretation pass required. That's a meaningful distinction for anyone dealing with financial regulators who want to know why a model recommended a mortgage over a credit card for a specific person, not just an aggregate feature-importance chart.
On the infrastructure side, nothing here is exotic—AWS Glue handles schema unification and service-category mapping via PySpark, SageMaker Processing jobs build time-windowed transaction features across 7 to 365-day windows using Dask, and everything lands as Snappy-compressed Parquet on S3 for the compression and column-pruning benefits. Training happens on ml.g5.12xlarge instances with four A10G GPUs, and the whole thing runs through SageMaker Pipelines with CloudWatch handling drift detection. AWS is explicit that this is an architecture overview, not a deployment tutorial, and they flag PII handling and least-privilege IAM as things you need to sort out yourself before touching real customer records.
What's notable is how much of this generalizes past banking. Multi-tower architectures with per-tower attention weighting are a reasonable pattern anywhere you've got heterogeneous data types feeding one prediction—insurance, retail, telecom churn models, you name it. The banking framing just happens to be where the explainability requirement is loudest, thanks to regulation rather than good intentions.
My take — AI-written commentary, not fact-checked reporting
I like that this treats explainability as an architectural decision rather than an afterthought—too much of the industry still bolts SHAP onto a black box and calls it compliance theater. The real tell here is that AWS is selling infrastructure and reference architecture, not a magic model; the GRU-over-LSTM and Glue-over-manual-ETL choices are the boring, correct engineering calls that never make headlines but are the actual difference between a demo and something that survives an audit.
Read more about this at: AWS