I am trying to understand how to compare KL divergence values. I have 3 sets of data, Good (G), Bad (B) and Unknown (U). I compute the symmetric form of KL for:
Sym_GU = KL(G||B) + KL(B||G) Sym_BU = KL(U||B) + KL(B||U)
Now, the classifier uses the following formulation:
if Sym_GU < Sym_BU: Good (lower KL values signify higher similarity)
Now, here is my question, say we get values like Sym_GU=.45 and Sym_BU=.48, do we still accept Good i.e. can we compute something like a p-value for KL or a way to encode confidence levels ?