A New Approach for Dimensionality Reduction and Variable Selection
Authors: Di Bo, Hoon Hwangbo, Vinit Sharma, Corey Arndt, Stephanie TerMaath; Published in Journal of Machine Learning Research (JMLR), 2023.
Abstract
High-dimensional data analysis provides a detailed understanding of systems but is often hindered by the curse of dimensionality. Existing dimensionality reduction techniques extract important features but lack interpretability and connection to decision making. Variable selection techniques offer interpretability but may fail to capture important interactions or require extensive computations. This research presents a novel method that identifies critical subspaces, which are reduced-dimensional physical spaces, for both dimensionality reduction and variable selection. The method utilizes a randomized search for subspace exploration and leverages ensemble techniques to improve model performance. When applied to high-dimensional data from the failure prediction of a composite/metal hybrid structure with complex progressive damage failure under loading, the proposed method outperforms existing alternatives in prediction and important variable selection.
[Abstract]