Shrunk Support Vector Clustering

Compared with Support Vector Machine (SVM) that has shown success in classification tasks, Support Vector Clustering (SVC) is not widely viewed as a competitor to popular clustering algorithms. The reason is easy to state that classical SVC is of high cost and moderate performance. In spite of ever-appearing variants of SVC, they fail in solving two problems well. Focusing on these two problems, this paper proposes a Shrunk Support Vector Clustering (SSVC) algorithm that makes an effort to address two difficulties simultaneously. In the optimization piece SSVC pursues a shrunk hypersphere in feature space that only dense-region data are included in. In the labeling piece of SSVC, a new labeling approach is designed to cluster support vectors firstly, and then label other data. The development of the shrunk hypersphere is implemented by optimizing a strongly convex objective, which can be converted to a linear equation system. A fast training method is given to reduce the heavy computation burden that is necessary in SVC to solve a quadratic optimization problem. The new labeling approach is based on geometric nature of the shrunk model and works in a simple but informed way. That removes the randomness encoded in SVC labeling piece and then improves clustering accuracy. Experiments indicate SSVC's better performance and efficiency than its peers and much appealing facility compared with the state of the art.