Heavy flavor identification at CMS with deep neural networks

Abstract

At the Large Hadron Collider, the identification of jets originating from heavy flavour quarks (b or c-tagging) is important for searches for new physics and for measurements of standard model processes. A variety of b-tagging algorithms has been developed by CMS to select b-quark jets based on variables such as the impact parameters of the charged-particle tracks, the properties of reconstructed decay vertices, and the presence or absence of a lepton, or combinations thereof. These algorithms heavily rely on machine learning tools and are thus natural candidates for advanced tools like deep neural networks. A new algorithm, DeepCSV, uses a deep neural network. The input is the same set of observables used by the existing CSVv2 b-tagger, with the extension that it uses information of more tracks. Also, the training strategy was adapted and about 50 million jets are used for the training of the deep neural network. The new DeepCSV algorithm outperforms the CSVv2 tagger, with an absolute b-tagging efficiency improvement of about 4% for a misidentification probability for light-flavour jets of 1%. In addition, DeepCSV is a multiclassifier simultaneously trained for c-tagging. For c-tagging DeepCSV outperforms the other taggers in CMS.

Glossary

CSVv2: Combined Secondary Vertex version 2 algorithm, based on secondary vertex and track-based lifetime informations, it is an updated version of the CSV algorithm used in Run 1 combining the variables with a neural network instead of a likelihood ratio and the secondary vertex information is obtained with the Inclusive Vertex Finder algorithm.

CSVv2L, CSVv2M, CSVv2T: CSVv2 algorithm at the loose, medium, tight operating points, defined as the values of the discriminator cut for which the rate for misidentifying a light jet as a b jet is 10%, 1%, and 0.1%, respectively.

DeepCSV: a new algorithm based on the same set of observables used by the CSVv2 b-tagger, with a simple extension to use more charged particle tracks. This algorithm is based on a deep neural network training, with four hidden layer (i.e. six layers altogether) of a width of 100 nodes each.

DeepCSVL, DeepCSVM, DeepCSVT: DeepCSV algorithm at the loose, medium, tight operating points, defined as the values of the discriminator cut for which the rate for misidentifying a light jet as a b jet is 10%, 1%, and 0.1%, respectively.

c-tagger: a c jet identification algorithm exploiting properties related to displaced tracks, secondary vertices, and soft leptons inside the jets. The training of the classifiers is performed using a Gradient Boosting Classifier (GBC). Two separate GBCs are provided, one for discriminating c jets from light jets (CvsL) and one for discriminating c jets from b jets (CvsB).

cMVAv2: combined MultiVariate Algorithm version 2, using a Boosted Decision Tree taking as input the different algorithm outputs of CSVv2, a variant of CSVv2 using another vertex reconstruction, Jet Probability (JP), Jet B Probability (JBP), Soft Electron (SE) and Soft Muon (SM) taggers.

JP: Jet Probability algorithm, based on the likelihood of tracks to come from the primary vertex (using the impact parameter significance values).

mu+jets: Measured b-tagging efficiency in multijet events with a muon, based on the combination of the results from different measurements, obtained using the PtRel, the LT and the System8 methods.

PtRel: Method for the measurement of the b-tagging efficiency in multijet events based on the transverse momenta of muons w.r.t. the jet axis.

System8: Method for the measurement of the b-tagging efficiency in multijet events with a muon, solving a system of 8 equations.

LT: Lifetime Tagging method for the measurement of the b-tagging efficiency in multijet events, based on template fits to the JP distributions.

Kin: Method for the measurement of the b-tagging efficiency in ttbar events in the dileptonic channel, based on a template fit to an MVA discriminator combining kinematic variables.

TagCount: Method for the measurement of the b-tagging efficiency in ttbar events in the dileptonic channel. The b-tagging efficiency is obtained by counting the number of events with two b-tagged jets in the selected sample of events.

TnP: Method for the measurement of the b-tagging efficiency in ttbar events in the semileptonic channel. The b-tagging efficiency is measured with a tag and probe method (TnP). As a tagging requirement, the CSVv2M requirement is applied to either the b-jet on the hadronic or leptonic side, while the b-jet from the other side is used as probe.

IterativeFit: Method for the measurement of the b-tagging efficiency in ttbar events in the dileptonic channel. This method is based on the calibration of the full b-tagging discriminator shape.

Simulation Plots at 13 TeV(click on plot to get the .pdf version )

For more details: Identification of b quark jets at the CMS Experiment in the LHC Run 2, CMS Collab., CMS PAS BTV-15-001.

For more details on the c-tagging algorithm see: Identification of c-quark jets at the CMS experiment, CMS Collab., CMS PAS BTV-16-001.

Roc curves

Figure Caption
perf_Log.pdf Figure 1: Performance of the b jet identification efficiency algorithms demonstrating the probability for non-b jets to be misidentified as b jet as a function of the efficiency to correctly identify b jets. The curves are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV , b jets from gluon splitting to a pair of b quarks are considered as b jets. The lines shown are for CSVv2, DeepCSV, and cMVAv2. cMVAv2 uses also the information from the soft leptons inside jets, while CSVv2, DeepCSV do not. The performance in this figure serves as an illustration since the b jet identification efficiency depends on the pT and η distribution of the jets in the topology as well as the amount of b jets from gluon splitting in the sample.

perf_cvsb_Log.pdf Figure 2: Performance of the c jet identification efficiency algorithms demonstrating the probability for b jets to be misidentified as c jet as a function of the efficiency to correctly identify c jets. The curves are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV , b jets from gluon splitting to a pair of b quarks are considered as b jets. The lines shown are for CSVv2, DeepCSV CvsB, c-tagger CvsB and cMVAv2. cMVAv2 and the c-tagger use also the information from the soft leptons inside jets, while CSVv2, DeepCSV do not.

perf_cvsl_Log.pdf Figure 3: Performance of the c jet identification efficiency algorithms demonstrating the probability for light jets to be misidentified as c jet as a function of the efficiency to correctly identify c jets. The curves are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV , b jets from gluon splitting to a pair of b quarks are considered as b jets. The lines shown are for CSVv2, DeepCSV CvsL, c-tagger CvsL and cMVAv2. cMVAv2 and the c-tagger use also the information from the soft leptons inside jets, while CSVv2, DeepCSV do not. The irregularity observed in the ROC curve of the c-tagger is caused by a sharp feature in the discriminator distribution due to jets without any selected tracks.

Figure Caption
effvspt_b_csvv2.pdf effvspt_b_deep.pdf Figure 4: b-jet efficiency as a function of the jet transverse momentum for the CSVv2 and DeepCSV algorithms. These efficiencies are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV. The last bin includes the overflow.

effvspt_c_csvv2.pdf effvspt_c_deep.pdf Figure 5: c-jet efficiency as a function of the jet transverse momentum for the CSVv2 and DeepCSV algorithms. These efficiencies are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV. The last bin includes the overflow.

effvspt_l_csvv2.pdf effvspt_l_deep.pdf Figure 6: light jet efficiency as a function of the jet transverse momentum for the CSVv2 and DeepCSV algorithms. These efficiencies are obtained on simulated ttbar events using jets within tracker acceptance with pT>30 GeV. The last bin includes the overflow.

Performance Plots in data at 13 TeV

Commissioning plots in AK4 jets

Figure Caption
CSVIVF_Log.pdf DeepCSV_Log.pdf Figure 7: Distribution of the CSVv2 (top) and DeepCSV (bottom) discriminators for ak4 jets in a muon enriched jet sample. The markers correspond to the data. The stacked, coloured histograms indicate the contributions of the different jet flavours in the simulation. Simulated events involving gluon splitting to b quarks (“b from gluon splitting”) are indicated separately from the other b quark production processes (“b”). The distributions from the simulation have been scaled to match the observed number of entries in data. The last bins of the histograms contain all entries above the histogram range. The underflow bin is included in the first bin.

Performance measurements in AK4 jets

Figure Caption
SFbCSVComp_Moriond17.pdf Figure 8: Comparison between the scale factors measured by different methods in ttbar events (Kin, TagCount, TnP, IterativeFit), the combined scale factors obtained from the muon enriched sample (mu+jets), and the combined scale factors obtained from ttbar and muon enriched samples (comb). The "comb" combined scale factors are based on the mu+jets, Kin and TnP measurements for CSVv2 and on the mu+jets and Kin measurements for DeepCSV. The scale factors measured in the muon enriched sample are averaged over the observed pT spectrum of the b jets from ttbar decays. For the IterativeFit method a cumulative scale factor for jets with pT above 30 GeV is extracted to allow a comparison.
Topic attachments
I Attachment History Action Size Date Who Comment
PDFpdf CSVIVF_Log.pdf r1 manage 24.8 K 2017-03-14 - 10:54 CarolineCollard  
PNGpng CSVIVF_Log.png r1 manage 19.8 K 2017-03-14 - 10:54 CarolineCollard  
PDFpdf DeepCSV_Log.pdf r1 manage 25.0 K 2017-03-14 - 10:54 CarolineCollard  
PNGpng DeepCSV_Log.png r1 manage 19.7 K 2017-03-14 - 10:54 CarolineCollard  
PDFpdf SFbCSVComp_Moriond17.pdf r1 manage 15.7 K 2017-03-14 - 10:54 CarolineCollard  
PNGpng SFbCSVComp_Moriond17.png r1 manage 35.9 K 2017-03-14 - 10:54 CarolineCollard  
PDFpdf effvspt_b_csvv2.pdf r1 manage 17.4 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_b_csvv2.png r1 manage 18.1 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf effvspt_b_deep.pdf r1 manage 17.5 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_b_deep.png r1 manage 18.2 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf effvspt_c_csvv2.pdf r1 manage 17.1 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_c_csvv2.png r1 manage 16.5 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf effvspt_c_deep.pdf r1 manage 17.0 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_c_deep.png r1 manage 16.7 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf effvspt_l_csvv2.pdf r1 manage 17.1 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_l_csvv2.png r1 manage 16.6 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf effvspt_l_deep.pdf r1 manage 17.2 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng effvspt_l_deep.png r1 manage 16.9 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf perf_Log.pdf r1 manage 54.1 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng perf_Log.png r1 manage 32.5 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf perf_cvsb_Log.pdf r1 manage 40.8 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng perf_cvsb_Log.png r1 manage 27.5 K 2017-03-14 - 10:50 CarolineCollard  
PDFpdf perf_cvsl_Log.pdf r1 manage 41.3 K 2017-03-14 - 10:50 CarolineCollard  
PNGpng perf_cvsl_Log.png r1 manage 28.7 K 2017-03-14 - 10:50 CarolineCollard  
Edit | Attach | Watch | Print version | History: r1 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r1 - 2017-03-14 - CarolineCollard
 
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    CMSPublic All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright &© 2008-2021 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback