On test-time active feature selection through tractable acquisition functions

Athresh Karanam, Sriraam Natarajan

Keywords: Active feature selection, Tractable probabilistic models, probabilistic circuits

TL;DR: We propose a novel framework for test-time active feature selection that harnesses the tractability and expressiveness of probabilistic circuits (PCs) to compute otherwise intractable acquisition functions.

Abstract: Real-life test-time decision making frequently necessitates the acquisition of additional features, which can often be costly. For instance, diagnosing a patient may require conducting laboratory tests or performing imaging scans. The acquisition of these potentially expensive features on a per-instance basis enables the efficient and effective allocation of resources. However, existing acquisition functions often pose computational challenges, necessitating the use of approximations. To address this issue, we introduce $MEASURES$, a framework for test-time active feature selection. Our approach harnesses the tractability and expressive efficiency of Probabilistic Circuits (PCs), a class of deep tractable probabilistic models, to compute otherwise intractable acquisition functions. Our experiments demonstrate the superior effectiveness of tractable acquisition functions compared to existing approaches.

Submission Number: 11