Dominant Set Based Density Kernel and Clustering

The density peak based clustering algorithm has been shown to be a potential clustering approach. The key of this approach is to isolate and identify cluster centers by estimating the local density of data appropriately. However, existing density kernels are usually dependent on user-specified parameters evidently. In order to eliminate the parameter dependence, in this paper we study the definition of dominant set, which is a graph-theoretic concept of a cluster. As a result, we find that the weights of data in a dominant set provides a non-parametric measure of data density. Based on this observation, we then present an algorithm to estimate data density without parameter input. Experiments on various datasets and comparison with other density kernels demonstrate the effectiveness of our algorithm.