Efficient Spectral Clustering using Core Points

Abstract

Spectral clustering is based on the graph Laplacian, and outputs good results for well-separated groups of points even when they have nonlinear boundaries. However, it is generally difficult to classify a large amount of data by this technique because computational complexity is large. We propose an algorithm using the concept of core points in DBSCAN. This algorithm first applies DBSCAN for core points and performs spectral clustering for each cluster obtained from the first step. Efficiency of the proposed algorithm is shown by the analysis of complexity. Simulation examples are used to show performance of the proposed algorithm.