Change LeptonID

The list of lines for which the NN cut is made available is the following:|

MuonNN	Cut	BW division	Electron NN	Cut	BW division
DiMuonDisplacedLine	0.1	Yes	Hlt1TrackElectronMVA	0.7	Yes
DiMuonHighMassLine	0.1	Yes	DisplacedDielectronLine	0.7	Yes
DiMuonNoIPLine	0.95	No	LowMassDielectronNoIPLine	0.95	Yes
DiMuonDrellYanLine	Chi2Corr cuts	No	LowMassDielectronDisplacedLine	0.8	Yes
TrackMuonMVA	0.1	Yes
Hlt1SMOG2DisplacedDiMuonLine	Chi2Corr	No
Hlt1SMOG2JpsiToMuMuTapLine	Chi2Corr	No
Hlt1SMOG2SingleMuon	Chi2Corr	No
DiMuonDisplacedSoftPTLine	0.8	No

(Then actually all lines cutting on isElectron are changed since the MR also changes the defition of isElectron to be NN based > 0.6)

I have set different muonNN cuts following the logic that was chosen to select Chi2Corr cuts. These cuts are a bit more aggressive than Chi2Corr but hopefully it will help recover more efficiency as they should allow for much softer pT cuts. Then for some specific lines some more stringent cuts are chosen, these are especifically the Dark Photon lines.

This MR showcases an example to change the LeptonID based on the NN operators:

• Electrons: Define isElectron based on the NN setting a cut similarly efficient but reducing background twice as much:

NN
MB:
Hlt1DiElectronDisplaced:            21/ 10000, (  423.00 +/-    35.37) kHz
jpsiee:
Hlt1DiElectronDisplaced:           2346// 10000, (10035.00 +/-   141.54) kHz


EoP
MB:
Hlt1DiElectronDisplaced:             47/ 10000, (  141.00 +/-    20.52) kHz
jpsiee:
Hlt1DiElectronDisplaced:           2291/ 10000, ( 6873.00 +/-   126.08) kHz

ROC curves (A bit shaky because the fits are a bit unstable...)

p_T \in [500,800] MeV	p_T \in [800,1200] MeV

p_T \in [1200,2500] MeV	p_T \in [2500,4000] MeV

p \in [3000,6000] MeV	p \in [6000,12000] MeV

p \in [12000,20000] MeV	p > 20000 MeV

Features data(top) simulation (bottom): (slight cut on the NN to cleanup the electron background)

EoP	EoP3x3	delta_barycenter

dispersion_x	dispersion_y	dispersion_xy

• Muons: Go line by line and modify the Chi2Corr cut > 1.8 to loose cut of 0.1 which improves the background reduction also by 50% at the same efficiency point. In this example I modified the DiMuonMassLine which shows these numbers for the Ks2MuMu as signal and MB as background. We see that we x10 lower rate on background, lowering the PT cut down to 0 MeV while keeping 50% more of the low PT muons.

NN
MB:
Hlt1DiMuonDisplaced:      9/ 10000, (   27.00 +/-     9.00) kHz
ks2mumu:
Hlt1DiMuonDisplaced:    222/ 10000, (  666.00 +/-    44.20) kHz

Chi2Corr and PT cut at 500 MeV
MB:
Hlt1DiMuonDisplaced:               91/ 10000, (  273.00 +/-    28.49) kHz
ks2mumu:
Hlt1DiMuonDisplaced:              147/ 10000, (  441.00 +/-    36.10) kHz

ROC curves (pion rejection:

p_T \in [100,500] MeV	p_T \in [500,1000] MeV	p_T \in [1000,2000] MeV

p \in [3000,7000] MeV	p \in [7000,12000 MeV	p > 12000 MeV

ROC curves (proton rejection, for this type of background the NN does not improve):

p_T \in [500,1000] MeV	p_T \in [1000,2000] MeV

p \in [3000,7000] MeV	p \in [7000,12000 MeV	p > 12000 MeV

Features data(top; muon vs proton)(mid; muon vs pion); simulation (bottom; actually trained on protons!):

chi2coor	chi2uncorr	dtx

dty	dty_velo
Goes with: lhcb-datapkg/ParamFiles!128 (merged) (NN models update)

Edit: I Have updated the roc curves, I'm refining the J/psi model, that's why they changed (are better).