We choosed Day6 16:00:00 to Day7 15:59:59 as the 1st day;
We choosed Day7 16:00:00 to Day8 15:59:59 as the 2nd day;
We choosed Day8 16:00:00 to Day9 15:59:59 as the 3rd day;
We added 8 hours to time, thus transformed a day from 16:00:00-15:59:59 into 0:00:00-23:59:59. As a result, we now have 3 complete days. Let's call them tr_7, tr_8, tr_9.
Then we used tr_8 for training, and tr_9 for validation, since we had limited RAM, and we would like to use some features like conversion rate.
After we trained a model with tr_8 for training, and tr_9 for validation, we finished tuning, and we got the best iteration round. Then, we used tr_8+tr_9 for training without any validation. We used the same parameters, and set the num_round to 1.1× the best iteration round we mentioned before. We would have a new model(train on tr_8+tr_9), and we used this model to predict the test data.

"Next click" was absolutely the magic feature! Besides, we used normal aggregate functions such as count or nunique.
In addtion, we used conversion rate, we calculated the rate on tr_7, and merged it on tr_8; calculated on tr_8, and merged it on tr_9; calculated on tr_9, and merged it on test.
Moreover, numpy can be very useful and way faster than pandas, especially in groupby functions.

We only used an LGB single model. We tried rank_avg with other models, but it didn't work... We could not figure out the reason.