A Novel Committee-Based Clustering Method

It is well recognized that clustering algorithms play an important role in data analysis. For a successful application of these algorithms, it is crucial to determine the relevant features in the original dataset. To deal with this problem there are efficient techniques for feature selection in the literature. Moreover, it is also well known that, in the clustering task, it is also difficult to define an adequate number of clusters. This paper proposes a new ensemble clustering method that is comprised of three stages: the first generates the clustering ensemble, the second combines the results of the multiple clustering scenarios generated, and the last one creates a new partition using the combined data. To generate the clustering ensemble, the method combines feature selection strategies and clustering with various numbers of clusters to produce a similarity matrix. This similarity matrix is then used to compute the final clustering output. Experiments performed using seven well known datasets showed the effectiveness of the proposed technique.