Constructing Rough Set Based Unbalanced Binary Tree for Feature Selection

Feature selection is one of the challenging problems facing data analysis in areas such as pattern recognition, data mining, and decision support. Many rough set algorithms for feature selection have been developed, most of which are essentially dependent on the definite information contained within the lower approximation. This paper proposes a novel approach, called Unbalanced binary tree based feature selection (UBT-FS), which utilizes the indefinite information contained within rough set boundary region for reduction. UBT-FS designs the underlying mechanism for obtaining the boundary region from the unbalanced binary tree and adopts the boundary region based significance for determining the optimal search path as well as the boundary region based evaluation criterion for identifying feature subsets. These allow UBT-FS to have considerable ability in finding an optimal or suboptimal reduct whilst simultaneously achieving obviously better computational efficiency than other available algorithms, which is also supported by the experimental results.