Building an explainable next-best-product recommendation system for banking on AWS
Amazon Web Services
This post presents a deep learning-based Next-Best-Product (NBP) recommendation system for banking, built using Amazon SageMaker AI and PyTorch. The architecture uses a multi-tower neural network with learned attention to provide per-customer explainability, addressing the challenge of predicting which product a customer needs next from diverse data sources.
The solution described on the AWS ML blog uses a multi-tower deep learning architecture with four specialized neural network towers processing different aspects of customer data: product adoption history, transactions, demographics, and behavioral segments. The towers are fused using a learned attention mechanism that provides per-customer explainability. The tech stack includes AWS Glue for ETL, PyTorch for the deep learning framework, Amazon S3 for data storage, and SageMaker AI for training and inference. The data pipeline involves two stages: an AWS Glue ETL job for data unification and an Amazon SageMaker Processing job for ML-specific feature engineering. The model architecture includes a Sequence Tower using a 2-layer GRU, Transaction, Customer, and Behavioral towers using MLPs, and a Feature Importance Module that produces per-customer importance scores. Training uses Adam optimizer with ReduceLROnPlateau scheduler, gradient clipping, and early stopping. The solution is designed to handle banking data with regulatory requirements for explainability.
- Сокращения
- AWS = Amazon Web Services — Amazon Web Services
- NBP = Next-Best-Product — следующий лучший продукт
- ETL = Extract, Transform, Load — извлечение, преобразование, загрузка
- ML = Machine Learning — машинное обучение
- GPU = Graphics Processing Unit — графический процессор
- IAM = Identity and Access Management — управление идентификацией и доступом
- GRU = Gated Recurrent Unit — гейтированная рекуррентная единица
- MLP = Multilayer Perceptron — многослойный перцептрон
- LSTM = Long Short-Term Memory — долгая краткосрочная память
- ReLU = Rectified Linear Unit — ReLU
- SHAP = SHapley Additive exPlanations — SHAP
- LIME = Local Interpretable Model-agnostic Explanations — LIME
- S3 = Simple Storage Service — Simple Storage Service
- DPU = Data Processing Unit — единица обработки данных
Source: AWS ML blog —
original