Attribute reduction based on directional semi-neighborhood rough set

Neighborhood rough set has become a mature means of attribute reduction for simplifying modeling and knowledge discovery in continuous and mixed data. It principally generates the neighborhood by condition attribute subset that allows monitoring the local class distribution, thereby evaluating the attribute significance. However, traditional neighborhood formed as a circle with a radius may not inspect some special local distributions correctly, especially near the class boundary. To alleviate such a problem, we present a directional semi-neighborhood formulation approach and redefine the corresponding neighborhood rough set for attribute reduction. Essentially, we attempt to form the neighborhood as a semi-circle directing to the homogeneous samples, which enables to reduce the negative influence of heterogeneous samples. Then, a novel neighborhood rough set using the directional semi-neighborhood is induced. Our feature evaluation criterion, i.e., neighborhood approximation quality established by upper approximation may be more satisfactory in reduct searching. Experiments validated on fifteen public benchmark data sets suggest the superiority of our proposal as it is better at selecting more informative condition attributes for building classification models.