Sequential Combination Methods for Data Clustering Analysis

This paper proposes the use of morethan one clustering method to improve clustering performance.Clustering is an optimization procedure based on a specific clusteringcriterion. Clustering combination can be regarded as a technique thatconstructs and processes multiple clustering criteria. Since the globaland local clustering criteria are complementary rather than competitive,combining these two types of clustering criteria may enhance theclustering performance. In our past work, a multi-objective programmingbased simultaneous clustering combination algorithm has been proposed,which incorporates multiple criteria into an objective function by aweighting method, and solves this problem with constrained nonlinearoptimization programming. But this algorithm has high computationalcomplexity. Here a sequential combination approach is investigated,which first uses the global criterion based clustering to produce an initialresult, then uses the local criterion based information to improve theinitial result with a probabilistic relaxation algorithm or linearadditive model. Compared with the simultaneous combination method,sequential combination has low computational complexity. Results on somesimulated data and standard test data are reported. It appears thatclustering performance improvement can be achieved at low cost throughsequential combination.