A Hierarchical Approach for Multi-task Logistic Regression

In the statistical pattern recognition field the number of samples to train a classifier is usually insufficient. Nevertheless, it has been shown that some learning domains can be divided in a set of related tasks, that can be simultaneously trained sharing information among the different tasks. This methodology is known as the multi-task learning paradigm. In this paper we propose a multi-task probabilistic logistic regression model and develop a learning algorithm based in this framework, which can deal with the small sample size problem. Our experiments performed in two independent databases from the UCI and a multi-task face classification experiment show the improved accuracies of the multi-task learning approach with respect to the single task approach when using the same probabilistic model.