Neighborhood Covering Based Rule Learning Algorithms

Abstract

At present, Rough set theory has been extensively discussed and studied in machine learning and data mining. Pawlak rough set theory provides a solid theoretical basis for processing incomplete or inconsistent data described by nominal attributes. However, it fails to deal with real classification tasks, in which most of the sample sets are numerical data. Recently, neighborhood covering methods based on rough sets demonstrate promising views in classification rule learning. These methods apply to numerical or complex data. In this paper, we put forward new neighborhood covering rule learning algorithms. We redefine the neighborhood radius to generalize the model, experiments produce good results in reducing the rule number, and more powerful generalization ability is expected. In terms of rule learning strategy, we combine kernel covering method, create an optimization classification hyperplane, and solve quadratic programming problem to train support coverings, like support vectors in SVM. Experiments show good classification performance of our improved algorithm.