A Robust Multi-Sphere SVC Algorithm Based on Parameter Estimation

To improve the robustness to noise, outliers and arbitrary cluster boundaries, a robust multi-sphere support vector clustering (SVC) algorithm is proposed in this paper. The proposed algorithm can automatically estimate a suitable kernel parameter, and determine the cluster number. The Gaussian kernel parameter is firstly estimated through a kernel parameter estimation algorithm which is based on support vector domain description (SVDD) and original local variance (LV) algorithm. Based on the estimated kernel parameter, the SVC algorithm classifies the given data points into different clusters and then the SVDD algorithm is performed several times for each cluster. At last, the membership is computed and the final clustering result is obtained based on these computed memberships. Several simulations verify the effectiveness of the proposed algorithm.