A reinforcement learning guided adaptive cost-sensitive feature acquisition method

Most of the existing feature selection methods tend to pursue the learning performance of the selected feature subset while ignoring the costs of acquiring each feature. However, in real-world problems, we often face the tradeoff between model performance and feature costs because of limited resources. Moreover, in some applications (e.g. medical tests), features are acquired sequentially in the learning process instead of having known the information of the whole feature set in advance. To solve these problems we design a reinforcement learning agent to guide the cost-sensitive feature acquisition process and propose a deep learning-based model to select the informative and lower-cost features for each instance adaptively. The whole process of feature acquisition will be determined by an agent according to what it has observed from inputs. In particular, a Recurrent Neural Network (RNN) model will learn the knowledge from the current sample and the agent will give the instructions on whether the RNN model will continue to select the next feature or stop the sequential feature acquisition process. Moreover, the proposed method can also select the features per block thus it can deal with high dimensional data. We evaluate the effectiveness of the proposed method on a variety of datasets including benchmark datasets, gene datasets, and medical datasets. Compared with the state-of-the-art feature selection methods, the proposed method can achieve comparable learning accuracy while maintaining lower feature costs. (C)& nbsp;2022 Elsevier B.V. All rights reserved.