Automated Feature Engineering with Reinforcement Learning

Abstract

Automatic Feature Engineering (AFE) aims to extract useful knowledge for interpretable predictions given data for the machine learning tasks of interest. Here, we develop AFE to extract dependency relationships that can be interpreted with functional formulas in order to discover physics meaning or new hypotheses for the problems of interest. We focus on materials science applications, where interpretable predictive modeling may provide a principled understanding of materials systems and guide new materials discovery. It is often computationally prohibitive to exhaust all the potential relationships to construct and search the whole feature space to identify interpretable and predictive features. We develop and evaluate new AFE strategies by exploring a feature generation tree with deep Q-network for scalable and efficient exploration policies. The developed reinforcement learning based AFE strategies are benchmarked with the existing AFE methods on several materials science datasets.