Dear All,
I hope this is not too off topic.
Apologies for not sending now any code, but the point is really for me
to understand how to proceed.
Let's say that you have a multiclass classification problem and the
outcome you want to predict is given by 9 different classes {A, B...}.
By training several models and checking the confusion matrix, I
noticed that it is particularly hard to tell apart class C and D.
What is the best way forward?
I am thinking about creating two artificial classes: C-and-D and then aggregate
all the other levels into a fictitious class "Other".
Then I can train a model to identify C-and-D vs the Other class.
The next step would be to train a model on telling apart only C and D
which I had merged in the C-and-D class.
Does such an approach sound sensible?
Many thanks
Lorenzo