Stochastic Encodings for Active Feature Acquisition

TL;DR

This paper introduces a stochastic latent variable model for active feature acquisition, enabling more reliable and non-myopic feature selection by reasoning over multiple possible feature realizations, outperforming existing methods.

Contribution

It proposes a novel supervised latent variable approach for active feature acquisition that overcomes limitations of reinforcement learning and greedy methods.

Findings

Outperforms diverse baselines on synthetic and real datasets

Provides reliable, non-myopic feature acquisition decisions

Demonstrates robustness across various data scenarios

Abstract

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.