Object-Neighbourhood Clustering Ensemble Method

Clustering is an unsupervised learning and clustering results are often inconsistent and unreliable when different clustering algorithms are used. In this paper we have proposed a clustering ensemble framework, named Object-Neighbourhood Clustering Ensemble (ONCE), to improve the consistency, reliability and quality of the clustering result. The core of the ONCE is a new consensus function that addresses the uncertain agreements between members by taking the neighbourhood relationship between object pairs into account in the similarity matrix. The experiments are carried out on 11 benchmark datasets. The results show that our ensemble method outperforms the co-association method, when the Average linkage is used. Furthermore, the results show that our ensemble method is more accurate than the baseline algorithm, and this indicates that the clustering ensemble method is more consistent and reliable than a single clustering algorithm.