Revealing the structure of prediction models through feature interaction detection

In recent years, machine learning models have been employed for prediction in various domains. While the prediction performance has obviously improved, some models have become too complex to understand, and these models are called black-box models. Detecting the feature interactions is a useful technique to gain insight into the structure of black-box models. In this paper, we propose a method based on high dimensional model representation (HDMR) to reveal the structure of prediction models by detecting the feature interactions that are embedded in the models. The HDMR-based method can detect the k-way interactions without any constraints on k and can detect the interactions from both classification and regression models. Moreover, this method is model-agnostic and can detect both global and local interactions. Experiments on some synthetic and real datasets demonstrate that the HDMR-based method can detect feature interactions effectively and improve the model interpretability. (C) 2021 Elsevier B.V. All rights reserved.