A Deeper Understanding of Active Feature Acquisition

Abstract

Data and decision making are increasingly becoming more prevalent. Reinforce ment learning (RL) has in recent years become an important field for tackling sequential decision making problems. One of the strengths of RL is its general framework of decision making which has been successfully applied to a wide range of problems. However, the common assumption of fully observable and available data in RL can be a limitation in problems where data is costly to acquire or missing entirely. Active Feature Acquisition (AFA) is a subfield of machine learning concerned with the former problem. Namely, to sequentially acquire features to maximize predictive performance under acquisition costs. However, despite rapid methodological progress, evaluation comparison of AFA methods remains difficult because methods are often evaluated under incompatible protocols. Further, most existing work assumes access to fully available data even though realworld datasets often contain missing values. Our contributions in this thesis are twofold. First, we present 𝙰𝙵𝙰𝙱𝚎𝚗𝚌𝚑, the first benchmarking framework for AFA methods in the classification setting. Along with it, we present the synthetic dataset C𝚄𝙱𝙴-𝙽𝙼, which is designed to evaluate potential tradeoffs between myopic and nonmyopic methods. Second, we extend the existing theoretical framework of the Partially Observable Markov Decision Process (POMDP) of the AFA problem to the missing data setting. We present novel theoretical insights that can be used to understand AFA under missing data. These incremental steps are important for deeper evaluation and understanding of AFA methods. We hope that our 𝙰𝙵𝙰𝙱𝚎𝚗𝚌𝚑 framework gets adopted by the community and that our theoretical insights can drive future research for foundational understanding of AFA under missing data.