Multi-objective optimization of ensemble of regression trees using genetic algorithms

We consider a prediction problem with multiple output responses based on an ensemble of multivariate regression trees. The selection of the optimal ensemble is formulated as a multi-objective optimization problem and solved using genetic algorithms. We illustrate the application of our approach on drug sensitivity prediction problem where the proposed methodology outperforms regular multivariate random forests in terms of correlation coefficients between predicted and experimental sensitivities. We also demonstrate that generating the Pareto-optimal front provides us a choice of ensembles for different optimization objectives.