In the classic feature selection task, the goal is to find a small number of features that are sufficient to separate two states. Thus, given two highly correlated features – the output will pick one and remove the other. However, in many applications, one would like to remain with all correlated features. For example, obtaining a complete subset of correlated genes in computational biology may help reveal biological mechanisms underlying medical states. In this project, we are developing data-driven approaches to the task of detecting groups of differential features.
Link to papers:
https://www.pnas.org/content/118/22/e2100293118.short
Link to code: