A new approach to exploring rough set boundary region for feature selection

Feature selection offers a crucial way to reduce the irrelevant and misleading features for a given problem, while retaining the underlying semantics of selected features. Whilst maintaining the quality of problem-solving (e.g., classification), a superior feature selection process should be reduce the number of attributes as much as possible. In this paper, a non-unique decision value (NDV), which is defined as the number of attribute values that can lead to non-unique decision values, is proposed to rapidly capture the uncertainty in the boundary region of a granular space. Also, as an evaluator of the selected feature subset, an NDV-based differentiation entropy (NDE) is introduced to implement a novel feature selection process. The experimental results demonstrate that the selected features by the proposed approach outperform those attained by other state-of-the-art feature selection methods, in respect of both the size of reduction and the classification accuracy.