CORE: Nonparametric Clustering of Large Numeric Databases

Current clustering techniques are able to identify arbitrarily shaped clusters in the presence of noise, but depend on carefully chosen model parameters. The choice of model parameters is difficult: it depends on the data and the clustering technique at hand, and finding good model parameters often requires time consuming human interaction. In this paper we propose CORE, a new nonparametric clustering technique that explicitly computes the local maxima of the density and represents them with cores. CORE proposes an adaptive grid and gradients to define and compute the cores of clusters. The incrementally constructed adaptive grid and the gradients make the identification of cores robust, scalable, and independent of small density fluctuations. Our experimental studies show that CORE without any carefully chosen model parameters produces better quality clustering than related techniques and is efficient for large datasets.