An Ensemble Density-based Clustering Method

Density based clustering is sound for its great ability of finding arbitrary shapes of clusters and identifying the number of clusters automatically. DBSCAN is a frequently used density based clustering algorithm. In DBSCAN a density threshold, which is hard to be chosen adaptively, should be specified to determine whether an object is dense or sparse. In this paper we introduce the concept of clustering ensemble to avoid the difficulty of selecting a single appropriate threshold. Performing DBSCAN multiple times with diverse thresholds picked up from a pre-constructed interval, the final partition can be figured out via a consensus function. Experimental results show that this method can go beyond DBSCAN both in validity and stability, and avoid the inefficiency caused by any inappropriate thresholds.