A new multi-task learning method with universum data

Abstract

Multi-task learning (MTL) obtains a better classifier than single-task learning (STL) by sharing information between tasks within the multi-task models. Most existing multi-task learning models only focus on the data of the target tasks during training, and ignore the data of non-target tasks that may be contained in the target tasks. In this way, Universum data can be added to classifier training as prior knowledge, and these data do not belong to any indicated categories. In this paper, we address the problem of multi-task learning with Universum data, which improves utilization of non-target task data. We introduce Universum learning to make non-target task data act as prior knowledge and propose a novel multi-task support vector machine with Universum data (U-MTLSVM). Based on the characteristics of MTL, each task have corresponding Universum data to provide prior knowledge. We then utilize the Lagrange method to solve the optimization problem so as to obtain the multi-task classifiers. Then, conduct experiments to compare the performance of the proposed method with several baslines on different data sets. The experimental results demonstrate the effectiveness of the proposed methods for multi-task classification.