Weighted ensemble of algorithms for complex data clustering

• Suggests to evaluate a “competence” of clustering algorithms in classifying complex data structures.
• Suggests a model of pairwise weighted clustering ensemble.
• Founds an expression for upper bound of error probability in partitioning an object pair into clusters.
• Proposes a mathematical methodology of determining optimal weights in ensembles of different clustering algorithms.
• Proposes Pairwise Weighted Ensemble Clustering algorithm (PWEC).

This paper considers a problem of clustering complex data composed from various structures. A collection of different algorithms is used for the analysis. The main idea is based on the assumption that each algorithm is “specialized” (as a rule, gives more accurate partition results) on particular types of structures. The degree of algorithm’s “competence” is determined by usage of weights attributed to each pair of observations. Optimal weights are specified by the analysis of partial ensemble solutions with use of the proposed model of clustering ensemble. The efficiency of the suggested approach is demonstrated with Monte-Carlo modeling.