This paper proposes learning-based approaches for transcoding MPEG-2 video into HEVC. In the training mode of the transcoder, mappings between extracted features and split decisions are calculated. While in the transcoding mode, the split decisions of coding units of the HEVC video are predicted. Two formulations are proposed for the prediction of split decisions based on multi model and multi-tier solutions. In the former solution, multi models are generated based on the total number of split flags in a coding unit. While in the latter solution, split decisions are modelled at three different coding depths. The proposed solutions are evaluated in terms of excessive bitrate, drop in PSNR, classification accuracy, model generation time and transcoding speedup. It is shown that the multi-tier solution maintains the rate-distortion behaviour of full re-encoding at the expense of lower gain in transcoding speedup. In comparison to existing work, it is shown that the proposed solutions offer a significant enhancement in terms of rate-distortion performance and classification accuracy.