Sequential combination methods for data clustering analysis

This paper proposes the use of more than one clustering method to improve clustering performance. Clustering is an optimization procedure based on a specific clustering criterion. Clustering combination can be regarded as a technique that constructs and processes multiple clustering criteria. Since the global and local clustering criteria are complementary rather than competitive, combining these two types of clustering criteria may enhance the clustering performance. In our past work, a multi-objective programming based simultaneous clustering combination algorithm has been proposed, which incorporates multiple criteria into an objective function by a weighting method, and solves this problem with constrained nonlinear optimization programming. But this algorithm has high computational complexity. Here a sequential combination approach is investigated, which first uses the global criterion based clustering to produce an initial result, then uses the local criterion based information to improve the initial result with a probabilistic relaxation algorithm or linear additive model. Compared with the simultaneous combination method, sequential combination has low computational complexity. Results on some simulated data and standard test data are reported. It appears that clustering performance improvement can be achieved at low cost through sequential combination.