Abstract
Density-based clustering is able to find clusters of arbitrary sizes and shapes while effectively separating noise. Despite its advantage over other types of clustering, it is well-known that most density-based algorithms face the same challenge of finding clusters with varied densities. Recently, ReScale, a principled density-ratio preprocessing technique, enables a density-based clustering algorithm to identify clusters with varied densities. However, because the technique is based on one-dimensional scaling, it does not do well in datasets which require multi-dimensional scaling. In this paper, we propose a multi-dimensional scaling method, named DScale, which rescales based on the computed distance. It overcomes the key weakness of ReScale and requires one less parameter while maintaining the simplicity of the implementation. Our empirical evaluation shows that DScale has better clustering performance than ReScale for three existing density-based algorithms, i.e., DBSCAN, OPTICS and DP, on synthetic and real-world datasets.