A game-theoretic rough set approach for handling missing data in clustering

Abstract

An important issue in clustering is how to deal with data containing missing values. A three-way decision making approach has recently been introduced for this purpose. It includes an added option of deferment which is exercised whenever it is not clear to include or exclude an object from a cluster. A critical issue in the three-way approach is how to decide the thresholds defining the three types of decisions. We examine the role of game-theoretic rough set model (GTRS) to address this issue. The GTRS model induces three-way decisions by implementing a game between multiple cooperative or competitive criteria. In particular, a game in GTRS is proposed which realizes the determination of thresholds from the viewpoint of tradeoff between accuracy and generality of clustering. Experimental results are reported for two datasets from UCI machine learning repository. The comparison of the GTRS results with another three-way model of (1, 0) suggests that the GTRS model significantly improves generality by upto 65% while maintaining similar levels of accuracy. In comparison to the (0.5, 0.5) model, the GTRS improves accuracy by upto 5% at a cost of some decrease in generality.