Validation-Driven Algorithm for SVC

This paper presents a novel approach, called validation-driven algorithm (VDA) for improving support vector clustering (SVC). The SVC algorithm is a well-known kernel-based clustering approach, but its clustering results keenly rely on the proper choice of parameters of kernel functions. Particularly, we have found that performance of SVC is very sensitive to the initial value of the kernel parameter qini, and a poor initial q value inevitably will incur heavy computation time for obtaining a satisfactory result. Furthermore, it is very difficult to determine the proper q value for input containing clusters with large diversity in variations. Previous SVC algorithms use an identical trial-anderror q for all clusters. Aiming to tackle these issues, VDA not only capable of finding a good initial value of the q parameter, but also provides a computationally efficient procedure for verifying the validity of the parameter setting as well as the clustering result. Simulation results demonstrate the effectiveness of the proposed VDA algorithm.