AWS skips SHAP, builds explanations into the model

AWS published an architecture for a bank product recommendation system that generates its own attribution scores instead of bolting on SHAP, per its engineering blog.

Banks that automate next-best-product pitches usually have to justify each one to a regulator after the fact. That normally means bolting on a separate tool like SHAP or LIME to explain the call, but AWS's design skips that step entirely.

Four towers feed one fusion layer. A two-layer GRU tracks product adoption history, while three MLP towers cover transactions, demographics, and behavioral segmentation.

The transactions tower runs across five windows: 7, 30, 60, 180, and 365 days. Each of the four towers outputs a 64-dimensional vector, and instead of concatenating them, the model fuses the vectors with multihead attention.

Attention weights fall out per-customer and per-tower by construction. That's why a recommendation can come with a breakdown like 40% product sequence, 30% transaction patterns, 20% demographics, and 10% behavioral segment. The number comes straight from the forward pass, not a separate run afterward.

The design generalizes past banking, per AWS. Any setting where a regulator is going to ask why a model made a call benefits from the same approach.

Design fusion for attributability and a post-hoc SHAP or LIME pass turns into insurance nobody needs to buy. Watch whether other regulated-industry recommendation systems, insurance, lending, healthcare, start describing their fusion layers instead of their accuracy metrics.

As of July 25, this reads as an architecture pattern for engineers building regulated recommendation systems to copy, not a new AWS product to buy.