A rough clustering algorithm for mining outliers in categorical data

Abstract

Outlier detection is an important data mining task with applications in various domains. Mining of outliers in data has to deal with uncertainty regarding the membership of such outlier objects to one of the normal groups (classes) of objects. In this context, a soft computing approach based on rough sets happens to be a better choice to handle such mining tasks. Motivated by this requirement, a novel rough clustering algorithm is proposed here by modifying the basic k -modes algorithm to incorporate the lower and upper approximation properties of rough sets. The proposed algorithm includes the necessary computational steps required for determining the object assignment to various clusters and the modified centroid (mode) computation on categorical data. An experimental evaluation of the proposed rough k -modes algorithm is also presented here to demonstrate its performance in detecting outliers using various benchmark categorical data sets. © Springer-Verlag 2013.