[Public Homepage] [Internal Homepage]
Neural Network b-Tagger

Neural Network b-Tagger for SecVtx-Tagged Jets
H.Bachacou, P.Lujan, M. McFarlane, W. Yao, LBNL


Network Formulation
This NN b-tagger reduces charm (c) and light (l) quark backgrounds while maintaining much of the bottom (b) signal. Its use of two networks allows flexibility and optimization for different physics. The version described on this page applies to jets tagged by SecVtx.

One of these networks is trained to separate b from l, the other, b from c. The networks' output distributions are below. The x-axes indicate the network assignments to the various species of quarks: 0 is background-like and 1 is signal-like; the black line represents b jets, the red line represents c jets and the blue line represents l jets.

The three maps below show jet efficiency (color) as a function of two cuts for tt Monte Carlo:

The network relies on the 16 variables below. The b-l network is trained with 1000 epochs, 10 hidden nodes and one output node while the b-c network is trained with 12 hidden nodes.
Variables from the SecVtx tagOther variables
Number of tracksReconstructed mass of Pass 1 tracks
Transverse decay length (L2D)Reconstructed mass of Pass 2 tracks
Transverse decay length Significance (L2D / σL2D)    Number of Pass 1 tracks
Chi SquaredNumber of Pass 2 tracks
Pseudo-cτ (=L2D × MSVX / pT, SVX)pT of Pass 1 tracks / Σ jet pT
Vertex MasspT of Pass 2 tracks / Σ jet pT
Vertex pT / Σ jet pTNumber of good tracks
Pass 1 or 2JetProb
Data and Monte Carlo Distributions

Systematics
The NN b-tagger scale factor (efficiency of b data ÷ efficiency of b Monte Carlo) is 0.97 ± 0.02. Since this is for SecVtx-tagged jets, this scale factor compounds with that of SecVtx (0.9xx ± 0.08). See data and Monte Carlo network outputs, input variable distributions and further treatment here.

Applications
• Application to the tt production cross-section. The NN b-tagger yields a measurement of the cross-section consistent with the same measurement performed with SecVtx only.
• Application to the Higgs search.

References
• Introductory discussion of neural networks
• JETNET 3.0 Manual