Comparative Analysis of Anti-clusters Formed Using Various Distance Metrics and k-Medoids Algorithm
Clusters, when formed, usually consist of similar data points. There are several cases where clusters should be formed by grouping dissimilar data points together. Clustering algorithms are not meant to group dissimilar data points. Hence in this paper, we propose a novel technique called anti-clustering which uses normal clustering techniques for grouping the data points but uses a novel distance matrix called complementary distance matrix for clustering dissimilar data points. We testify to this technique using various distance metrics such as Euclidean, Manhattan, Chebyshev, Canberra and Minkowski. We also compare the quality of anti-clusters, obtained by using these distance metrics, using the Silhouette coefficient. Through experimental evaluations, we analyze the results and prove that anti-clustering is practical and can be implemented efficiently.