I have an unbalanced binary problem where i'm interested in optimizing the recall of the 1% of the points with the highest score given by the algorithm itself. I'm using a lightgbm with these feval MeanChosen and fobj MeanChosen_loss to see if I get it, but I don't know why the results are really bad. Does anyone know what might be going on or how I could achieve it?
The only thing I need is to accumulate the maximum number of elements of class 1 possible in the 1% of the data where the algorithm gives me a higher score.
Code of the feval and fobj used
I tried at first with standard feval and fobj and it worked fine. When i first included only this new customized feval te results improved, but now that i included the fobj it started giving really bad results.
