Neighborhood density information in clustering

Density Based Clustering (DBC) methods are capable of identifying arbitrary shaped data clusters in the presence of noise. DBC methods are based on the notion of local neighborhood density estimation. A major drawback of DBC methods is their poor performance in high-dimensions. In this work, a novel DBC method that performs well in high-dimensions is presented. The novelty of the proposed method can be summed up as follows: a hybrid first-second order optimization algorithm for identifying high-density data points; an adaptive scan radius for identifying reachable points. Theoretical results on the validity of the proposed method are presented in this work. The effectiveness and efficiency of the proposed approach are illustrated via rigorous experimental evaluations. The proposed method is compared with the well known DBC methods on synthetic and real data from the literature. Both internal and external cluster validation measures are used to evaluate the performance of the proposed method.