Consensus clustering algorithm based on the automatic partitioning similarity graph

Consensus clustering has been recently applied as a solution to the clustering problem. This combines multiple clusterings of a set of objects into a single integrated clustering. Consensus clustering algorithms attempt to find stable and robust results by composing calculated results from the base clustering algorithms. However, finding a consensus clustering algorithm capable of obtaining the final clusters automatically has remained a challenge. Furthermore, most of them are affected by an outlier. In the present paper, a cluster-based consensus clustering algorithm is proposed based on partitioning similarity graph in which each vertex is a cluster composed of a set of points. Within the proposed algorithm, the Cosine, Jaccard, and Dice similarity measures are used to measure the similarity between two vertices. The proposed algorithm operates in three steps. First, outlier clusters are automatically obtained by pruning similarity graph. Then, meta-clusters are obtained by splitting graph and merging sub-graphs. Finally, based on the meta-clusters, the consensus solution and outlier points are obtained using majority voting. The proposed algorithm has linear time complexity in terms of the number of points. The number of clusters also is obtained automatically in a consensus solution. To evaluate the performance of the algorithm, real and artificial datasets were used. The obtained results showed a dramatic improvement in the accuracy of the final clusters.