try clustering of features only present in either test/train to compute frequencies

[aranas]

We might wanna give this a try:
https://www.kaggle.com/nanomathias/feature-engineering-importance-testing/comments#302507