A new multi-prototype based clustering algorithm

K-means is a well-known prototype based clustering algorithm for its simplicity and efficiency. However, most k-means methods assume different classes are represented by one prototype, which makes a limit of k-means algorithms. Recently, multi-prototype clustering methods have been raised to tackle this problem, which composed of two stages: split stage and merge stage. For multi-prototype algorithms, a proper prototype number plays a vital role in the algorithm performance and it is generally given by users in a trial and error way. In this paper, a new incremental k-means clustering algorithm is designed to determine the appropriate prototype number automatically. Firstly, a new indicator is presented to judge whether the number of prototype is appropriate in the split stage. Secondly, a new merge indicator is defined according to the distance formula from datapoint to hyperplane in the merge stage. Finally, simulation results on 8 datasets illustrate the effectiveness and superiority of the proposed algorithm.