Active Feature Selection for the Mutual Information Criterion

We study active feature selection, a novel selection setting in which unlabeled data is available, but the budget for labels limited, and examples to label can be actively selected by algorithm. focus on using classical mutual information criterion, selects k features with largest label. In setting, goal use significantly fewer than set size still find whose based entire large. explain experimentally choices that we make algorithm, show they lead successful compared other more naive approaches. Our design draws insights relate problem of pure-exploration multi-armed bandits settings. While here information, our general methodology adapted feature-quality measures as well. The extended version this paper, reporting all experiment results, available at Schnapp Sabato (2020). code following url: https://github.com/ShacharSchnapp/ActiveFeatureSelection