AWS publishes a multi-tower recommender that builds explainability into the forward pass instead of bolting on SHAP

AWS Machine Learning Blog·medium signal

AWS details a banking next-best-product system using four specialized towers — a 2-layer GRU sequence tower over product adoption history, plus MLP towers for transactions (7/30/60/180/365-day windows), customer demographics, and behavioral segmentation — each emitting a 64-dim vector fused by multihead attention. Because fusion is attention-based rather than concatenation, per-customer tower weights fall out natively, producing attributions like '40% product sequence, 30% transaction patterns, 20% demographics, 10% behavioral segment.' The architectural argument generalizes beyond banking: designing for interpretability beats post-hoc SHAP or LIME when regulators demand transparency.