Metric-Guided Multi-task Learning

Abstract

Multi-task learning (MTL) aims to solve multiple related learning tasks simultaneously so that the useful information in one specific task can be utilized by other tasks in order to improve the learning performance of all tasks. Many representative MTL methods have been proposed to characterize the relationship between different learning tasks. However, the existing methods have not explicitly quantified the distance or similarity of different tasks, which is actually of great importance in modeling the task relation for MTL. In this paper, we propose a novel method called Metric-guided MTL which explicitly measures the task distance using a metric learning strategy. Specifically, we measure the distance between different tasks using their projection parameters and learn a distance metric accordingly, so that the similar tasks are close to each other while the uncorrelated tasks are faraway from each other, in terms of the learned distance metric. With a metric-guided regularizer incorporated in the proposed objective function, we open a new way to explore the related information among tasks. The proposed method can be efficiently solved via an alternative method. Experiments on both synthetic and real-world benchmark datasets demonstrate the superiority of the proposed method over existing MTL methods in terms of prediction accuracy.