Convex formulation for multi-task L1-, L2-, and LS-SVMs

Quite often a machine learning problem lends itself to be split in several well-defined subproblems, or tasks. The goal of Multi-Task Learning (MTL) is leverage the joint from two different perspectives: on one hand, single, overall model, and other hand task-specific models. In this way, found solution by MTL may better than those either common Starting with work Evgeniou et al., support vector machines (SVMs) have lent themselves naturally approach. This paper proposes convex formulation for L1-, L2- LS-SVM models that results dual problems quite similar single-task ones, but multi-task kernels; turn, makes possible train using standard solvers. As an alternative approach, direct optimal combination already trained can also considered. paper, procedure compute combining parameter respect four error functions derived. shown experimentally, proposed approach performs generally combination, both them are straight use