An Approach to Investigate the Effectiveness of Multi-Task Learning

Abstract

Multi-task learning has been widely used in different applications where it is required to build a single model that optimizes many associated learning tasks at the same time. Thus, in this modeling task, the same set of features is used by a single model that is trained and optimized to produce simultaneously more than one output. In this paper, an approach for investigating the relationship between the features and modeled outputs is presented. First, the set of relevant features to each individual output is generated. Then, a different set of features is used depending on the correlation between the outputs. The used features include the intersection, union, and difference between features relevant to each output individually. The performance of the multitask-learning model is evaluated and compared to the performance of the corresponding single-task model in each case. Obtained results indicate that if the outputs to be detected are correlated, then the best performance of a multi-task model is obtained, in the case where the intersection of relevant features to each output is used. On the other hand, in the case where the outputs are not correlated, the union of the set of relevant features to each individual output produces the best results. Moreover, building multi-task models for uncorrelated outputs gives worse results in all cases when compared to the single-task model. Thus, obtained results shed the light on when it is more viable to build multi-task learning models as compared to single task models. In addition, it presents some guidelines on the features to be used based on the correlations between the outputs to be detected by the multi-task models.