Transfer Learning via Auxiliary Labels with Application to Cold-Hardiness Prediction