Active Feature Acquisition is an instance-wise, sequential decision making problem. The aim is to dynamically select which feature to measure based on current observations, independently for each test instance. Common approaches either use Reinforcement Learning, which experiences training difficulties, or greedily maximize the conditional mutual information of the label and unobserved features, which makes myopic acquisitions. To address these shortcomings, we introduce a latent variable model, trained in a supervised manner. Acquisitions are made by reasoning about the features across many possible unobserved realizations in a stochastic latent space. Extensive evaluation on a large range of synthetic and real datasets demonstrates that our approach reliably outperforms a diverse set of baselines.

Problem: In the real world not all information to solve a problem is all available at once. An AI model should be able to look at its current information and decide what to measure next to improve its prediction. Doctors do this when they diagnose a patient: based on their existing observations they choose what test to conduct next.Solution: We developed an AI-based method that addresses this problem. It makes decisions similarly to how a human might. It considers many possible situations and which measurement tends to be the best.Impact: With more development this method could be used as a diagnostic tool.