Efficient, interpretable and automated feature engineering for bank data

Abstract

Banks rely on expert-generated features and simple models to have high performance and interpretability at the same time. Interpretability is needed for internal assessment and regulatory compliance for specific problems such as risk assessment and both expert generated features and simple models satisfy this need. However, feature generation by experts is a time-consuming process and susceptible to bias. In addition, features need to be generated fairly often due to the dynamic nature of bank data, and in case of significant changes or new data sources, expertise might take a while to build up. Complex models, such as deep neural networks, may be able to remedy this. However, interpretability/explainability approaches for complex models are not satisfactory from the banks' point of view. In addition, such models do not always work well with tabular data which is abundant in banking applications. This paper introduces an automated feature synthesis pipeline that creates informative and domain-interpretable features which iconsumes significantly less time than brute-force methods. We create novel feature synthesis steps, define elimination rules to rule out uninterpretable features, and combine performance-based feature selection methods to pick desirable ones to build our models. Our results on two different datasets show that the features generated with our pipeline; (1) perform on par or better than features generated by existing methods, (2) are obtained faster, and (3) are domain-interpretable.