A Fast and Stable Cluster Labeling Method for Support Vector Clustering

Abstract

Even though support vector clustering (SVC) is able to handle arbitrary cluster shapes effectively, its popularity is frequently degraded by highly intensive time complexity, poor label performance and even instability for efficiency. To overcome such problems, a fast and stable cluster labeling (FSCL) method is proposed. Based on stable equilibrium points, the FSCL first finds an appropriate division of support vectors. With a nonlinear sample sequence strategy presented here, the connected components profiled by support vectors (SVs) can be determined in terms of sampling all stable equilibrium point pairs; and the FSCL prefers a density centroid constructed by one subset of SVs, along with a stable equilibrium point, to represent a component while avoiding local optimization. Finally, the remaining data points can be assigned the label of the nearest components with respect to a weighted distance. Time complexity analysis and comparative experiments suggest that the FSCL improves both the efficiency and clustering quality significantly while guaranteeing stability.