Forward backward greedy algorithms for multi-task learning with faster rates
A large body of algorithms have been proposed for multi-task learning. However, the effectiveness of many multi-task learning algorithms highly depends on the structural regularization, which incurs bias in the resulting estimators and leads to slower convergence rate. In this paper, we aim at developing a multi-task learning algorithm with faster convergence rate. In particular, we propose a general estimator for multitask learning with row sparsity constraint on the parameter matrix, i.e., the number of nonzero rows in the parameter matrix being small. The proposed estimator is a nonconvex optimization problem. In order to solve it, we develop a forward backward greedy algorithm with provable guarantee. More specifically, we prove that the output of the greedy algorithm attains a sharper estimation error bound than many state-of-the-art multi-task learning methods. Moreover, our estimator enjoys model selection consistency under a mild condition. Thorough experiments on both synthetic and real-world data demonstrate the effectiveness of our method and back up our theory.
