A hybrid model of ordinal ranking-based clustering using G+Rank K-Means

Abstract

K-Means is a clustering technique that maps object features onto multidimensional coordinates and groups them based on location closeness. However, measuring closest distance can be doubtful when ranking representation of ordinal scale objects are not considered. Thus, distribution of objects in a cluster may violate ranking representation. For example, a same-rank object may be grouped into different clusters. To address this issue, an enhanced of K-Means algorithm is proposed to achieve better and meaningful result of ranking-based clustering. It is based on integration of ranking algorithm that sort objects into ranking list which also representing object closeness based on its nearby location. A new additional step in K-Means is proposed in reassigning unaligned K-Means nearest objects using ranking attribute that eventually accelerates the clustering process. AHP ranking algorithm is integrated into K-Means in achieving a ranking-based cluster. This enhancement was evaluated on three ordinal datasets covering 67 Java programs, 92 students’ marks on computer architecture subject and 456 EUFA’s football club coefficient ranking list. The results show that by integrating ranking algorithm in K-Means as proposed in G+Rank K-Means, a rank cluster representation has been successfully achieved. The purity value that represents the correctness against certain group classification has also increased.