A Feature Reduction Approach Based on Adaptively Weighted Neighborhood Rough Sets

Feature reduction based on weighted neighborhood rough sets has been widely applied in machine learning and data mining. In existing weighted neighborhood rough set models, the weight of each feature is calculated independently and fixed without adjusting according to the selected feature set. Therefore, it cannot guarantee that the most significant feature for the selected feature set can always be selected. For this issue, we introduce adaptive weights into neighborhood relations, then propose a novel feature reduction approach. First, the separability of individual features and feature subsets is analyzed to adaptively calculate the weights of features based on the current selected feature set. Then, adaptively weighted neighborhood relations are defined to construct an adaptively weighted neighborhood rough set (AWNRS) model. Furthermore, a greedy search algorithm (DAWNRS) based on AWNRS is proposed to select a feature subset that has both high dependency and strong correlation. Finally, the proposed algorithm is compared with other representative reduction algorithms by using eight UCI datasets on three classifiers. Experimental results show that the proposed approach achieves superior classification performance.