Comparative density peaks clustering

Abstract Clustering analysis is one of the major topics in unsupervised machine learning. A recent study proposes a novel density-based clustering algorithm called the Density Peaks. It is based on two intuitive assumptions: that cluster centers have a higher density than those of their neighbors, and that they also have a relatively large distance from other points with a higher density. To see whether a distance is relatively large, we should make a comparison of it and another one. However, such comparison is not explicitly modeled in the algorithm. Therefore, we propose the Comparative Density Peaks algorithm which takes the comparison into the design of the method. Furthermore, we give our analysis of Density Peaks from the perspective of the tree structure, and summarize two sufficient conditions that contribute to a good clustering performance under the Density Peaks framework. Extensive experiments show that our proposed algorithm significantly outperforms the original Density Peaks clustering algorithm.