Utilization of Dynamic Reducts to Improve Performance of the Rule-Based Similarity Model for Highly-Dimensional Data

Abstract

This paper presents an extension to the Rule-Based Similarity (RBS) model -- a novel rough set approach to the problem of learning a similarity relation from data. The original model, proposed in [1], applied the notion of Tversky's feature contrast model in a rough set framework to facilitate an accurate case-based classification. In the dynamic RBS model, a dynamic reducts technique is used to broaden the scope of the considered similarity aspects. This is especially important when dealing with objects described by numerous attributes. The extended model was tested on several microarray datasets from RSCTC'2010 Discovery Challenge. The results proved that it is significantly more accurate than the original RBS as well as some other popular classification algorithms, such as the \emph{random forest} or $k$-NN combined with several attribute selection methods.