I don't know why we don't have a 10th position in the lb and so I'm assuming it's me 😅.

Anyways hello Zindians,

What a competition this was, I started this competition soon as it was launched and I tried my very best, I hoped for a better position but God's plans...

How I approached this problem from the beginning was just as it was stated, prev rows kpis, one row per file(totalling 7256 train rows, 1932 test rows). Never changed this approach at all , since I was pretty sure that was the objective. At the beginning the guys who we started the competition with, we Struggled to even get to 0.70 after multiple submissions and when new guys started streaming in and hacking 0.70 I knew there was something I was overlooking.

I only focused on the KPIS at the start and so just when I introduced the dates as features , I jumped to 0.70 welll, this showed that time had an impact on the fault rate as well and so I kept it as a feature. After that it was a matter of trying to get to 0.71 0.72 when Krishna join3d and eventually 0.75 , well I only achieved 0.71 after all that struggle,...that translated well in the private lb but was not good enough to take me to the top positions

My model is pretty simple, I dropped relation and fault duration, because I wondered how they told us to use them as features and at the time before the fault occurrence, we don't have that information? I stand to be corrected why we were told to use them as features , I also added lags , for the prev 2 and 3 days for only the data rate and calculated the mean of the data rate of all rows before the fault occurred. I never used any data at the actual time of fault occurence.

The only trick that maybe not many people utilized was thresholding, and that also had a huge impact on the scores.

At the end I used a mode ensemble of lightgbm , catboost and xgboost, funny enough this was inferior to one single untuned catboost 😂

Anyways a lot of bla bla bla,

Github Link: Please star the repo if you find it helpful

https://github.com/koleshjr/FAULT-IMPACT-ANALYSIS--HUAWEI-PROJECT/tree/main

Funny enough you can also share notebooks on ZIndi all you have to do is:

To embed your notebooks in discussions you need to;

1. Upload your notebook to GitHub Gist
2. Copy the unique URL from the Gist
3. Create a new discussion on Zindi and insert a one-liner in your discussion using this format:

::GIST::/[your username]/[Gist ID]

1. Where the text after ::GIST:: is the unique URL from the Gist