Ensembles of extremely randomized trees for multi-target regression

In this work, we address the task of learning ensembles of predictive models for predicting multiple continuous variables, i.e., multi-target regression (MTR). In contrast to standard regression, where the output is a single scalar value, in MTR the output is a data structure – a tuple/vector of continuous variables. The task of MTR is recently gaining increasing interest by the research community due to its applicability in a practically relevant domains. More specifically, we consider the EXTRA-TREE ensembles – the overall top performer in the DREAM4 and DREAM5 challenges for gene network reconstruction. We extend this method for the task of multi-target regression and call the extension EXTRA-PCTs ensembles. As base predictive models, we propose to use predictive clustering trees (PCTs) – a generalization of decision trees for predicting structured outputs, including multiple continuous variables. We consider both global and local prediction of the multiple variables, the former based on a single model that predicts all of the target variables simultaneously and the latter based on a collection of models, each predicting a single target variable. We conduct an experimental evaluation of the proposed method on a collection of 10 benchmark datasets for with multiple continuous targets and compare its performance to random forests of PCTs. The results reveal that a multi-target EXTRA-PCTs ensemble performs statistically significantly better than a single multi-target or single-target PCT. Next, the performance among the different ensemble learning methods is not statistically significantly different, while multi-target EXTRA-PCTs ensembles are the best performing method. Finally, in terms of efficiency (running times and model complexity), both multi-target variants of the ensemble methods are more efficient and produce smaller models as compared to the single-target ensembles.