Prototype Propagation Clustering Based on Large Margin

Feijiang Li, Yuhua Qian, Jieting Wang

Abstract: The idea of the prototype is important in clustering analysis. The main idea of prototype clustering is representing clusters by the compact model in the form of a set of prototypes and using the prototypes to guide the assignment of instances in the data set. Based on the idea of prototype clustering, a number of clustering algorithms have been proposed. However, most of the prototype-based algorithms focus on discovering prototypes, while the assignment step is simply handled by assigning each instance to its nearest prototype. Due to the low representability for the instances being far away from any prototypes, the nearest neighbor method may bring many mistakes. To handle this problem, in this paper, we draw large margin theory into the assignment step and propose a prototype propagation clustering algorithm. which gradually assigns the instances to the clusters. During the assignment process, the prototypes, clusters and the margin of instances are dynamically updated to guarantee sufficient representability of an instance's prototype. Visible experiments on four two-dimensional data sets are conducted to show the working mechanism of the proposed assignment method. Experimental analyses on benchmark data sets illustrate the superior clustering performance of the proposed clustering algorithm comparing with other five representative algorithms.