A Spectral Clustering Algorithm Based on Self-Adaption
Kan Li, Yushu Liu · 2007
In traditional spectral clustering algorithms, the number of cluster is choosen in advance. A self-adaption spectral clustering algorithm is proposed to decide the cluster number automatically, which eliminates the drawbacks of two kinds of spectral clustering methods. In our algorithm, eigengap is used to discover the clustering stability and decide the cluster number automatically. We prove theoretically the rationality of cluster number using matrix perturbation theory. A kernel based fuzzy c-means is introduced to spectral clustering algorithm to separate clusters. Finally the experiments prove that our algorithm tested in the UCI data sets may get better results than c-means, Ng et.al's algorithm and Francesco et.al's algorithm.