Add an attribute to the learning models to capture training-testing split of the dataset

[amoghskulkarni]

This is different in case of timeseries data since it cannot be sampled completely randomly. Some background knowledge about this is required.