Abstract: Clustering is to find the best partition of unlabeled observations under a certain group stucture hypothesis.Given the group stucturc hypothesis, the most clustering algorithms is to to iteratively optimize of fittness of data distribution (called algorithm validity). In fact, the clustering validity is determined by three factors: hypothesis, algorithm and apriori validity. Therefore, a variation of gaussian mixture model was proposed in this paper, then the measurement and estimation method of cluster structure stability were defined. Based on them, the SMCIus algorithm was designed to achieve the stable clustering structure by means of split merge operations. The experiment shows SMCIus' performance in clustering quality.

Key words: Clustering, Variation of mixture model, Structure robustness, Split Merge