stable!

0.8 wow!

UPDATE : i did a small code bug when creating labels,

Now CV : 0.7 LB : 0.7

what's your cv/lb looking like?

I tried different ways to generate training set...and surprisingly there is no consisitency in the scores. It's varying with great difference. I personally think that generating proper training set that is representative of the objective is key here.

Things I tried:

- Include all the rows from all files

- Include only rows where the fault > 0

- Include only the 1st instance when the fault > 0 / file

Any new approaches are most wellcome.

@koleshjr exactly like yours! It seems that we need to improve that hehe

@Rakesh_Jarupula yes, that's also what I'm experiencing. I think two points are essencial: 1) the way you set up the training set, and 2) how you input the NaNs on test set.

@yanteixeira Yeah I have been seen you following me closely and I think we have the same approach. I'm shocked by guys like @ff and @reacher getting 0.9.. cv and 0.8 cv respectively, I haven't encountered a cv of greater than the 0.65 - 0.69 region. Also , what I got from @AntonioDeDomenico is that we are supposed to be having one row per network element, if that is wrong I stand to be corrected.

I am referring to this discussion here:

https://zindi.africa/competitions/fault-impact-analysis-towards-service-oriented-network-operation-maintenance/discussions/17921

@AntonioDeDomenico Hi Rakesh, I hope i understand well your question. In the training set you need to label the datarate change when the fault occurs, max 1label per file, comparing the datarate in the row prior to the fault and the datarate measured when the fault appears.

I will review my approach.

Your are not wrong. We are supposed to be having one row per NE.

Same issue here! My model doesnt seem to learn anything. Most probabilites are around 0.5. Except some data_rate == 0 rows which are easy to classify as 0.

@Charrada Right now, you have an LB of 0.69. How can you say your model isn't learning? hahaha

@Koleshjr I'm also confused about how people are getting those high scores. Suddenly, the first position has an LB of 0.71... I think we should rethink our steps and try a different approach.

@yanteixeira that's trueee, but we seem to have a correlating cv vs lb , as the old competitive machine learning adage says: always trust you cv 😅

@yanteixeira a trick gave me a small boost. But still, my model is not performing well when looking at the predictions distribution. Maybe i wont use that submission for the private leaderboard :D

@ff It seems you and @Yisakberhanu have the same approach that is for the: CV = 0.63 & LB = 0.54

I think you are the first one with LB > CV

Very stable score

amazing score!