Rectified Decision Trees: Towards Interpretability, Compression and Empirical Soundness