An improved cohesion self-merging clustering algorithm

Hybrid clustering algorithms have long been focused on in machine learning research community. The most important components of this kind of algorithm are the split and merge criteria. The cohesion-based self-merging is an interesting hybrid clustering algorithm proposed in the literature. Although the algorithm has a good performance, it suffers from instability of its results. To alleviate the instability and make it more efficient, in this paper, we improve the two components of it. For the split component, an optimal method of determining initial prototypes for K-means is presented, for the merge component, the cohesion criterion is improved. The experimental results demonstrate the efficiency of the proposed method.