The final version of our paper entitled “Physics-constrained Automatic Feature Engineering for Predictive Modeling in Materials Science” recently presented at the 35th AAAI Conference on Artificial Intelligence (AAAI-21) is now accessible at the link below:
Ziyu Xiang, Mingzhou Fan, Guillermo Vázquez Tovar, William Trehem, Byung-Jun Yoon, Xiaofeng Qian, Raymundo Arroyave, Xiaoning Qian, “Physics-constrained Automatic Feature Engineering for Predictive Modeling in Materials Science“, 35th AAAI Conference on artificial intelligence (AAAI-21), Feb. 2-9, 2021.
Automatic Feature Engineering (AFE) aims to extract useful knowledge for interpretable predictions given data for the machine learning tasks of interest. In this work, we presented a novel AFE scheme that effectively extracts relationships from data that can be interpreted based on functional formulas to discover their “physical meaning” or “new hypotheses”. Here we focused on materials science applications, where interpretable predictive modeling may enhance our understanding of materials systems and also guide the discovery of new materials.
Typically, it is computationally prohibitive to exhaustively explore all potential relationships to identify interpretable and predictive features. We overcome this challenge by designing an AFE strategy that efficiently explores a feature generation tree (FGT) using a deep Q-network (DQN) for scalable and efficient exploration of the feature space in an automated manner. We demonstrate that our proposed DQN-based AFE strategy yields promising results when benchmarked against existing AFE methods based on several materials science datasets.