dsmilab / wsdm2018-kkbox-mrc Goto Github PK

Before tuning, need to explore all columns given by data.

Since considering features separately struggles so far, we need to discard the original data structure.

We don't have time fall into useless results which have been confirmed before.

Using only one thread to process large files is really slow.

When I tried to add the feature where the song which how many users have listened to, local CV score rose up; however, LB score fell down.

I didn't figure out why this happened.

I don't know how useful it is. But we need a try.