A novel feature selection method with transition similarity measure using reinforcement learning

Abstract

Feature selection identifies the relevant features and removes the irrelevant and redundant ones, intending to obtain the best-performing feature subset. This paper proposes a new feedback feature selection system with a reinforcement learning-based method to identify the best feature subset. The proposed system mainly includes three parts. First, decision tree branches are used to traverse the state space to discover new rules and select the best feature subset. Second, a transition similarity measure is introduced to ensure that the system keeps exploring the state space by creating diverse branches to overcome the redundancy problem. Finally, the informative features are the most involved in constructing the best branches. The performance of the proposed approaches is evaluated on nine standard benchmark datasets. The results in terms of AUC score, accuracy, and running time demonstrate the effectiveness of the proposed system, as it selects the fewest number of relevant features in less computational time.