A New Clustering Algorithm Based on Rewarding Support Vectors
Ling Ping, Rong Xiangsheng · 2019
Support vector exhibits impressive behaviors in supervised learning tasks like classification and regression. However, they are embarrassed in unsupervised learning like clustering, due to the fact that support vectors produced by unsupervised learning process cannot provide sufficient information that facilitates label detection. Existing clustering algorithms based on support vector are often pay high cost to detect clusters but present moderate and unstable results. For that, this paper extends the geometric meaning of support vectors from describing cluster contours to describing both cluster contours and inner-cluster structures. That is achieved by adopting the individually-tuned scale parameters in support vector computation. Based on that, the spectrum analysis is done to find the number of clusters and grouping schemas. Finally, data are labelled. Such a process formulates a new Clustering algorithm based on Rewarding Support Vectors (CRSV). Empirical evidence on benchmark datasets and real dataset demonstrates the improvement of CRSV over traditional support vector clustering approach as well as the peer variants, and the competitiveness with some state of the art clustering methods.