lol. xD

😂😂

Hello @AntonioDeDomenico 👋🏽,

Did you confirm this step finally?

@JuliusFx how's your cv/lb correlation looking like?

Hi @Koleshjr. I don't know whether to trust what I am having. Looks odd at 0.7 cv VS 0.72 lb. As you say, we keep building. About 7 more days to explore better solutions

Wait the last time I checked you had 0.73 cv 0.72 lb does this mean you are at 0.76cv??😲

no, the CV now is not correlated with LB.

CV: 0.74, LB: 0.75

CV: 0.75, LB: 0.74

That is why the concern.

Oh okay but damn that's still super impressive , 9 more days to figure out the trick 😅

There'll definitely be a shake up here. I got scared when I saw the distribution of my best LB . The ratio of 0/1 is nearly 25/75, which seems abnormal for "anomaly detection"-like task. Because you'd expect the opposite (having more 0s)

Yes, exactly.

That is due to the metric f1-score

I am little confused....How you got 25/75 ratio? For me it's balanced.

Amazing, and how does that correspond to the LB? @Rakesh_Jarupula i.e cv vs lb?

Yeah for submissions at LB 0.69 (or under), it is balanced. But over 0.70, I tend to lose the balance. ( Don't overfit the LB :) )

I believe it to be same for most of the guys above 0.7

Maybe the Top5 can relate to it.

I can relate to this completely. Subs above 0.70 for me are very random. Very sensitive to the number of folds you use as well, no stability whatsoever, but I believe some of the guys at >0.70 like @Krishna_Priya are not overfitting the lb, 0.73 cv 0.72 lb at first and now 0.74 cv 0.75 lb, thats stable very stable so Maybe there is just one thing that we are missing that he is capitalizing on, anyways we keep building

Actually I was taking about the Train set labels, Not Test set prediction distribution.

Test: Same 25/75

CV is around 0.7, But my LB score is not at all stable. If I make any small change, CV is reasonable but LB score changes drastically.

@Koleshjr, yeah I agree. Anyone with both CV/LB above 0.72 and very close, like Krishna, "should" be safe. Let's make sure to find that one thing 😅

@Rakesh_Jarupula I also see that instability when I try to overtune the threshold on my CV. So, I keep it more conservative, hence the stability and the gap between CV & LB

This is very true Tuo.