Clustering Similarity Comparison Using Density Profiles

Abstract. The unsupervised nature of cluster analysis means that objects can be clustered in many different ways. This means that different clustering al-gorithms can lead to vastly different results. To address this, clustering sim-ilarity comparison methods have traditionally been used to quantify the de-gree of similarity between alternative clusterings. However, existing tech-niques utilize only the point-to-cluster memberships to calculate the similar-ity, which can lead to unintuitive results. They also can’t be applied to analyze clusterings which only partially share points, which can be the case in stream clustering. In this paper we introduce a new measure named ADCO, which takes into account density profiles for each attribute and aims to address these problems. We provide experiments to demonstrate this new measure can of-ten provide a more reasonable similarity comparison between different clus-terings than existing methods.