Abstract:K-Means is the most popular clustering algorithm with the convergence to one of numerous local minima, which results in much sensitivity to initial representatives. Many researches are made to overcome the sensitivity of K-Means algorithm. However, this paper proposes a novel clustering algorithm called K-MeanSCAN by means of the local optimality and sensitivity of K-Means. The core idea is to build the connectivity between sub-clusters based on the multiple clustering results of K-Means, where these clustering results are distinct because of local optimality and sensitivity of K-Means. Then a weighted connected graph of the sub-clusters is constructed using the connectivity, and the sub-clusters are merged by the graph search algorithm. Theoretic analysis and experimental demonstrations show that K-MeanSCAN outperforms existing algorithms in clustering quality and efficiency.