Feature Interaction Selection for High-Dimensional Experimental Data

In a material development process, discerning the effect of material properties and their interactions on material behaviors is critical to achieving the desired functionality of a material. This causal analysis often involves a small experimental dataset arranged in a high dimension and is challenged by the curse of dimensionality. Feature selection can alleviate such a challenge by producing a short list of features that are significant, but identifying significant feature interactions is very challenging. In this proposal, we propose a couple of approaches that can evaluate and determine important interactions, including a randomized subspace-based model (RSM), feature subspace selection (FSS), and XGBoost. The core idea of RSM and FSS is to randomly generate low-dimensional subsets of a feature set and use such a subset, referred to as a subspace, as a basic modeling and significance evaluation unit. The significance of each potential subspace is evaluated, and a significant subspace is integrated into a base model to form an ensemble predicting the response. The experimental results also provide valuable insights into the damage process of the hybrid material structure. We upgrade the subspace generation and exploration by using XGBoost. The subspace is evaluated by its contribution to the response. When applied to the analysis of a composite/metal hybrid structure exhibiting complex progressive damage failure under loading, our methods show advantages in response modeling and feature selection compared to other machine learning-based alternatives.