Dealing with relative similarity in clustering: An indiscernibility based approach

Abstract

In this paper we propose a clustering method that works on relative proximity. The key process of this method is iterative refinement of N binary classifications, where N denotes the number of objects. First, for each of N objects, an equivalence relation that classifies all the other objects into two classes, similar and dissimilar, is assigned by refering to their relative proximity. Next, for each pair objects, we count the number of binary classifications in which the pair is included in the same class. We call this number as indiscernibility degree. If indiscernibility degree of the pair is larger than a user-defined threshold value, we modify the equivalence relations so that all of them commonly classify the pair into the same class. This process is repeated until clusters become stable. Consequently we get the clusters that follows granularity of the given threshold without using geometric measures.