SciPost Submission Page
Learning Selection Cuts With Gradients
by Mike Hance, Juan Robles
Submission summary
| Authors (as registered SciPost users): | Mike Hance |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2502.08615v2 (pdf) |
| Date accepted: | Oct. 23, 2025 |
| Date submitted: | Aug. 25, 2025, 6:13 p.m. |
| Submitted by: | Mike Hance |
| Submitted to: | SciPost Physics Core |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Experimental, Computational |
Abstract
Many analyses in high-energy physics rely on selection thresholds (cuts) applied to detector, particle, or event properties. Initial cut values can often be guessed from physical intuition, but cut optimization, especially for multiple features, is commonly performed by hand, or skipped entirely in favor of multivariate algorithms like BDTs or neural networks. We revisit this problem, and develop a cut optimization approach based on gradient descent. Cut thresholds are learned as parameters of a network with a simple architecture, and can be tuned to achieve a target signal efficiency through the use of custom loss functions. Contractive terms in the loss can be used to ensure a smooth evolution of cuts as functions of efficiency, particle kinematics, or event features. The method is used to classify events in a search for Supersymmetry, and the performance is compared with common classification tools. An implementation of this approach is available in a public code repository and python package.
Author comments upon resubmission
Thanks to you and the reviewers for your care in reading and making suggestions on our manuscript. The reviewer comments motivated some changes to the paper, which we summarize below; along with detailed notes addressing reviewer comments provided in separate responses.
Thanks again, and best wishes,
Juan and Mike
List of changes
With the addition of the BDT and the kCuts comparisons, we felt it better to focus on the comparisons of CABIN with those methods, evaluated in the form of summary ROC curves and cut evolution, rather than showing ROC curves based on CABIN output scores for individual efficiency targets (which were included in the original draft). This has simplified the discussion of the results and focused the attention on the results of CABIN’s regularization procedure.
Finally, we have taken this opportunity to do some language editing, cleanup of the bibliography, and adding some references. In particular, we have now included citations of recent LHC papers that use selection cuts, including a search for SUSY that motivated the use of the SUSY dataset under study.
Published as SciPost Phys. Core 8, 079 (2025)
Reports on this Submission
Report #2 by Wouter Verkerke (Referee 2) on 2025-10-1 (Invited Report)
Report
Requested changes
Figure 7 shows a rather striking difference in cut stability vs efficiency for TMVA and CABIN (with the latter being more stable), you observe this also in your results section, but a short sentence on why this is the case would be a nice addition (if possible).
Recommendation
Publish (meets expectations and criteria for this Journal)
Report
Recommendation
Publish (meets expectations and criteria for this Journal)
Author: Michael Hance on 2025-10-13 [id 5920]
(in reply to Report 2 by Wouter Verkerke on 2025-10-01)Thank you for your comments! We have added a short sentence in the paragraph on the plot of the loss values, noting that the "smoothness" term in the loss is responsible for the stability of the CABIN cuts compared to TMVA.