Clustering Ensembles Based on Normalized Edges

Abstract

Abstract. The co-association (CA) matrix was previously introduced to combine multiple partitions. In this paper, we analyze the CA matrix, and address its difference from the similarity matrix using Euclidean distance. We also explore how to find a proper and better algorithm to obtain the final partition using the CA matrix. To get more robust and reasonable clustering ensemble results, a new hierarchical clustering algorithm is proposed by developing a novel concept of normalized edges to measure the similarity between clusters. The experimental results of the proposed approach are compared with those of some single runs of well-known clustering algorithms and other ensemble methods and the comparison clearly demonstrates the effectiveness of our algorithm.

Enhanced Ensemble Clustering via Fast Propagation of Cluster-Wise Similarities

“…Ensemble clustering aims to combine a set of multiple base clusterings into a better and more robust consensus clustering result [20]. In the past decade, many ensemble clustering algorithms have been proposed [24], [25], [26], [29], [30], [31], [35], [36], [37], [38], [39], [40], [41], [42], [43], [44], which can be classified into three main categories, namely, the pair-wise co-occurrence based algorithms [20], [38], [41], the graph partitioning based algorithms [35], [36], [40], and the median partition based algorithms [24], [37], [39], [42].…”

“…The pair-wise co-occurrence based algorithms [20], [38], [41] typically build a co-association matrix by considering the frequency that two objects occur in the same cluster among the multiple base clusterings. By treating the co-association matrix as the similarity matrix, the hierarchical agglomerative clustering algorithms [19] can be used to obtain the consensus result.…”

“…Fred and Jain [20] for the first time presented the concept of co-association matrix and designed the evidence accumulation clustering (EAC) method. Then, Li et al [38] extended the EAC method by presenting a new hierarchical agglomerative clustering algorithm that takes the sizes of clusters into consideration via normalized edges. Yi et al [41] dealt with the uncertain entries in the co-association matrix by exploiting the matrix completion technique to improve the robustness of the consensus clustering.…”

“…Despite the fact that significant progress has been made in the ensemble clustering research in recent years [24], [25], [26], [35], [36], [37], [38], [39], [40], [41], [42], [47], [48], [49], [50], there are still two challenging issues in most of the existing algorithms. First, they mostly investigate the ensemble information at the object-level, but often fail to go beyond the object-level to explore the information at higher levels of granularity in the ensemble.…”

Robust Ensemble Clustering by Matrix Completion

“…The basic idea is to first find the corresponding cluster labels between different partitions, and then obtain the consensus partition through a voting process. The second group of approaches in this category is based on co-association/similarity matrix [13], [25], [41]. They use the similarity measure to combine multiple partitions, thus avoiding the label correspondence problem.…”