The problem of learning several related tasks has recently been addressed with success by the so-called multi-task formulation, that discovers underlying common structure between tasks. Metric Learning for Kernel Regression (MLKR) aims at finding the optimal linear subspace for reducing the squared error of a Nadaraya-Watson estimator. In this paper, we propose two Multi-Task extensions of MLKR. The first one is a direct application of multi-task formulation to MLKR algorithm and the second one, the so-called Hard-MT-MLKR, lets us learn same-complexity predictors with fewer parameters, reducing overfitting issues. We apply the proposed method to Action Unit (AU) intensity prediction as a response to the Facial Expression Recognition and Analysis challenge (FERA'15). Our system improves the baseline results on the test set by 24% in terms of Intraclass Correlation Coefficient (ICC).