Benchmarking multi-target regression methods

Machine learning methods for multi-target regression (MTR) rely on the hypothesis that an inter-target correlation can improve predictive performance. In the last years, many MTR methods were developed, but there are still questions about how their performances are influenced by the datasets characteristics such as linearity, number of targets, and inter-correlation complexity. Aiming at contributing to the understanding of the relationship between the dataset properties and MTR methods, we generated 33 synthetic datasets with controlled characteristics and tested their performance with single-target and six MTR methods. The results showed that MTR methods were able to improve performance even in datasets whose targets were not linearly correlated among them, but the predictive improvement differed among the combinations of method/regressor according to the dataset composition.