BEST DP twiki


Beyond Standard Model (BSM) signatures frequently decay with hadronic final states. At the CMS experiment large jets are clustered with the Anti-kT algorithm and 𝚫R = 0.8 to contain jets from boosted heavy objects. Jets from heavy objects have unique shapes that can be identified after a Lorentz boost along the jet axis to the center of mass frame. In this new rest frame, the particle candidates are re-clustered into AK4 jets and this information is used to calculate a series of event shape variables. The Boosted Event Shape Tagger (BEST) previously tagged jets in 2016 data from the CMS experiment. It classifies jets as Top, Higgs, Z, W, b, and QCD. BEST has been updated to a Full Run 2 tagger featuring two main improvements: an increase in statistics and new input features.

  • 280,000 training events per category 🠪 4,500,000 training events per category
  • New input features: p and E of the re-clustered jets, angles between re-clustered jets, and better b-tagging inputs
  • 271 input features

During the past three years, we also created a method for using BEST as an image tagger. Normally, BEST is trained on high level variables and the hope was that these images would provide low level information. However, this showed no improvement. We present our method to demonstrate the power of using a simple, dense neural network with physics inspired input features.

Public Plots

These plots were used for a poster at BOOST 2022.

Plot Caption
BEST figure1.png
Figure 1: A confusion matrix of the improved version of BEST using an independent test set of 2017 Run-2 Monte Carlo Simulated data. There is an overall improvement in classification accuracy with the most notable improvements being the bottom and QCD categories. The improvement in bottom and QCD is mostly due to the availability of better b-tagging at CMS
BEST figure2.png
Figure 2: A confusion matrix of the 2016 BEST architecture evaluated on 2017 Run-2 Monte Carlo Simulated data. Note that this has an overall worse accuracy than the improved version of BEST.
BEST figure3.png
Figure 3: A ROC curve for QCD comparing the old BEST architecture to the new one. All networks were evaluated on 2017 Run-2 Monte Carlo Simulated Data. A true positive is defined as any QCD jet correctly identified as QCD and a false positive is defined as any other jet (top, W, Z, Higgs, bottom) that was misidentified as QCD. Here we see a considerable improvement in BEST’s ability to tag QCD jets.
BEST figure4.png
Figure 4: An averaged ROC curve comparing the old BEST architecture to the new one. All networks were evaluated on 2017 Run-2 Monte Carlo Simulated Data. These averages were calculated using the machine learning approach of “macro” averaging for a multi-classifier
BEST figure5.png
Figure 5: Averaged images of AK8 Jets coming from W-jets boosted into different rest frames and displayed with the Cassini projection. The heat represents the energy of the candidates inside the bin. A useful way of thinking of these projections is to place the top of the positive Z-axis in the center of the top circle and the negative Z-axis in the center of the bottom circle. These images were used to train image-based versions of BEST. None of these image versions could perform better than our simple, dense network. A simple network with physics inspired inputs is a powerful tool.

-- BrendanRegnery - 2022-08-15

Topic attachments
I Attachment History Action Size Date Who Comment
PNGpng BEST_figure1.png r1 manage 81.5 K 2022-08-16 - 14:26 BrendanRegnery  
PNGpng BEST_figure2.png r1 manage 92.1 K 2022-08-16 - 14:26 BrendanRegnery  
PNGpng BEST_figure3.png r1 manage 59.6 K 2022-08-16 - 14:26 BrendanRegnery  
PNGpng BEST_figure4.png r1 manage 57.9 K 2022-08-16 - 14:26 BrendanRegnery  
PNGpng BEST_figure5.png r1 manage 115.6 K 2022-08-16 - 14:26 BrendanRegnery  
Edit | Attach | Watch | Print version | History: r3 < r2 < r1 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r3 - 2022-08-16 - BrendanRegnery
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    Sandbox All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright &© 2008-2022 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