Hello Zindians,
First I want to apologize for taking too long to publish my solution(The evaluation took too long to be finalized). The link to the github repo is attached below. Please star it as it motivates me to open source more of my work. It sucks to open source a high performing notebook and not get appreciated thus making me sometimes rethink my decision of open sourcing my top solutions. Thank you. Lastly @Krishna_Priya if you could open source your work as well I would really appreciate it since you were very competitive during the public lb phase and there is much I would love to learn from you . PLus other top performing guys you could as well open source your work as it helps the community to grow. Thanks in Advance.
Here is my solution:
https://www.kaggle.com/code/krishnapriya18/zindi-beam-traffic-ml
Public Rank 1, Private is way below :)
The same solution without a silly mistake lands private rank 2. There is an ensemble notebook on top of this which is just mean predictions of lgbm and catboost.
My mistake - forgot to remove 3 features from the model towards the end. ('thptime_lag_168','util_lag_168','usercount_lag_168')I appreciate it a lot . Thank you
Thanks for sharing your solution. Our team went with a pure NN approach with log transformed targets. 19 public -> 104 private. Obviously things went wrong. Will learn from your solutions and also continue sharing my own.
thanks @nymfree Most probably it was the cv and how you handled predictions for the last week after the time gap