A Method for Boundary Processing in Three-Way Decisions Based on Hierarchical Feature Representation

Abstract

For binary classification problem, all samples can be divided into three regions based on the three-way decision theory: positive regions, negative regions and boundary regions. These samples in boundary regions may be impossible to make a definite decision for lacking of detailed information. More information obtained from positive and negative regions is crucial for boundary processing. In the real word, people may identify positive regions based on one rule, and identify negative regions on another. The samples in boundary regions are also divided to positive or negative regions based on different rules. In this paper, we propose a method for processing boundary regions in three-way decisions based on hierarchical feature representation (\(HFR-TWD\)), which can obtain hierarchical feature representation of positive and negative regions. Firstly, all samples are divided into three regions by MinCA, which builds the most accurate covers for each class. Then samples in positive regions and negative regions respectively construct hierarchical feature representation. Thirdly, the best feature representation of each class is selected by using boundary region validating. Finally, boundary samples in test set are divided according to best feature representation of each class. Experiments show that the proposed method \(HFR-TWD\) improves classification accuracy.