Using associators to generate ensemble biclustering from multiple evolved biclusterings

Biclustering is a data mining technique that performs clustering of the rows and columns of a matrix simultaneously. An associator is a numerical measure of how closely associated two objects should be. Ensemble methods integrate information from multiple solutions to generate superior solutions. A simple evolutionary algorithm to quickly locate multiple biclusterings of synthetic test data. The good submatrices of these biclusterings are then used as associators. Associators are accumulated across many runs of the evolutionary algorithm to create a master association matrix. This matrix is then used, via simultaneous hierarchical clustering, to create a final ensemble biclustering. Results are presenting on tuning the evolutionary algorithm as well as for the overall biclustering algorithm. The algorithm correctly locates planted clusters in the data, providing proof of concept for the ensemble technique. The technique is modular with the evolutionary algorithm, fitness function, and ensemble integration technique all easily swapped for other techniques.