Risk/benefit analysis of including MCORR in the training features
Quick summary:
- By definition, the inclusion of the corrected mass in the training set might result in a TOPO response that ultimately sculpts the corrected mass shape. This is highly undesirable, and probably observed in Run 1 and Run 2 (Greg mentioned the low-mass in the
B_s \to K \mu \nu
analysis). - Preliminary results suggest a benefit in bolstering the classification power of the NN. This must be re-evaluated, keeping in mind efficiencies too.
Likely scenario: remove MCORR entirely OR hack the loss to mitigate the correlation with the response. The latter is probs painful and overkill, as I (Blaise) do not expect to see much of a change in efficiency and ROC score with and without MCORR - certainly not high enough to warrant the inclusion in the training features, at the expense of a systematic downstream in the analysis stage.