DiFA: Differentiable Feature Acquisition

Feature acquisition in predictive modeling is an important task many practical applications. For example, patient health prediction, we do not fully observe their personal features and need to dynamically select acquire. Our goal acquire a small subset of that maximize prediction performance. Recently, some works reformulated feature as Markov decision process applied reinforcement learning (RL) algorithms, where the reward reflects both performance cost. However, RL algorithms only use zeroth-order information on reward, which leads slow empirical convergence, especially when there are actions (number features) consider. modeling, it possible first-order i.e., gradients, since often given already collected dataset. Therefore, propose differentiable (DiFA), uses representation selection policy enable gradients flow from loss parameters. We conduct extensive experiments various real-world datasets show DiFA significantly outperforms existing methods number large.