The aLS-SVM based multi-task learning classifiers

The multi-task learning support vector machines (SVMs) have recently attracted considerable attention since the conventional single task learning ones usually ignore the relatedness among multiple related tasks and train them separately. Different from the single task learning, the multi-task learning methods can capture the correlation among tasks and achieve an improved performance by training all tasks simultaneously. In this paper, we make two assumptions on the relatedness among tasks. One is that the normal vectors of the related tasks share a certain common parameter value; the other is that the models of the related tasks are close enough and share a common model. Under these assumptions, we propose two multi-task learning methods, named as MTL-aLS-SVM I and MTL-aLS-SVM II respectively, for binary classification by taking full advantages of multi-task learning and the asymmetric least squared loss. MTL-aLS-SVM I seeks for a trade-off between the maximal expectile distance for each task model and the closeness of each task model to the averaged model. MTL-aLS-SVM II can use different kernel functions for different tasks, and it is an extension of the MTL-aLS-SVM I. Both of them can be easily implemented by solving quadratic programming. In addition, we develop their special cases which include L2-SVM based multi-task learning methods (MTL-L2-SVM I and MTL-L2-SVM II) and the least squares SVM (LS-SVM) based multi-task learning methods (MTL-LS-SVM I and MTL-LS-SVM II). Although the MTL-L2-SVM II and MTL-LS-SVM II appear in the form of special cases, they are firstly proposed in this paper. The experimental results show that the proposed methods are very encouraging.