K Modes Clustering Algorithm Based on a New Distance Measure

T he leading par tit ional clustering technique, K Modes, is one of the most computationally eff icient clustering methods fo r categ orical data. In the t raditional K Modes algo rithm, the simple matching dissim ilarity measure is used to compute the distance betw een two values of the same catego rical at t ributes. T his compares tw o categorical v alues directly and results in either a difference of zero w hen the tw o values are ident ical or one if otherw ise. How ever, the similarity betw een catego rical values is not considered. In this paper, a new distance measure based on rough set theory is pr opo sed, w hich overcomes the shortage of the simple matching dissimilarity measure and is used along w ith the t radit ional K Modes cluster ing algo rithm. While comput ing the distance betw een tw o values of the same catego rical att ributes, the new distance measure takes into account not only their dif ference but also discernibility o f other relat ional categ orical at tr ibutes to them. T he t ime complexity of the modif ied K Modes clustering alg orithm is linear w ith r espect to the number of data objects w hich can be applied for lar ge data sets. The performance of the K Modes alg orithm w ith the new distance measure is tested on real w orld data sets. Comparisons w ith the K Modes algor ithm based on many dif ferent distance measures illust rate the effect iveness o f the new distance measure.