LG: A clustering framework supported by point proximity relations

Abstract Clustering is a research problem based on the data's proximity relationship which is not made full use of by all the existing algorithms. In this paper, we present a novel two-stage LG framework consisting of the proposed Local Energy Gradient Oppression (LEGO) and the Guide Point Assignation (GPA) strategies which are closely related to the data points’ proximity relations. In the LG framework, it is crucial to locate the appropriate centers for the subsequent data label assignment, and therefore we introduce the nuclear model viewing the dataset as a collection of charged particles, which is the basis of LEGO, and the points with local maximum potential energy are ascertained as the cluster centers. Besides, the GPA strategy innovatively adopts the idea that the cluster center actively selects data points as the same cluster, enabling the LG framework still to be effective when dealing with datasets of arbitrary shape distribution. Superiorities of the proposed framework and the two strategies are demonstrated on four synthetic datasets and three real-world faces image datasets in terms of two clustering performance metrics.