apologies if this has been answered somewhere else.
As far as I am aware, the TMVA Likelihood method(s) Reader interface gives access only to the per-event likelihood ratio (L_s /( L_s+L_b)) in the application stage.
Is there a way to get the actual per-event signal (bkg) likelihood value L_s (L_b) instead?
I am interested in combining information from different independent sub-detectors in a global likelihood-based classifier, hence the need to first get the actual likelihoods for s(b) in each sub-detector.
Unfortunately this seems impossible in the current implementation, the only piece of information you can extract is the likliehood ratio.
One would have hoped that the individual likelihoods would be reconstructible using the multiclass version, but it seems the system of equations is ill-formed.
Now, implementing such as this is definitely possible, but a good argument would be needed so that it’s work could be prioritized.