Event Selection Strategies: From Cuts to Classifiers

August 17, 2026 · 5 min read

Why Event Selection Deserves a Second Look

If you have spent time on an analysis, you know that event selection is not a preamble — it is the analysis. Every signal-to-background ratio, every systematic uncertainty, every limit you set downstream depends on which events you kept and which you threw away. Most physicists learn selection through cut flows, and cut flows work. But as analyses push toward rarer signals and noisier environments, it is worth asking whether a classifier could do the same job better, and when that trade-off is actually justified.

This article walks through both approaches side by side so you can make an informed choice for your own analysis.

The Cut Flow: What It Does Well

A sequential cut flow applies a series of threshold requirements on observables — transverse momentum above some value, missing energy above another, a lepton isolation requirement, and so on. Each cut is interpretable: you can draw the distribution, mark the threshold, and explain the acceptance to a referee in two sentences.

Cut flows have real advantages:

• Transparency. Every selection step is auditable. A collaborator can reproduce your selection table from a single page of requirements.
• Systematic control. You can propagate uncertainties cut by cut and identify which requirement drives your dominant systematic.
• Stability. A cut on a well-modelled variable is unlikely to behave differently between data and simulation in a way you cannot diagnose.

The weakness is dimensionality. A cut flow selects a rectangular region in feature space — no matter how many variables you include, the boundaries are always axis-aligned hyperplanes. If the signal and background populations are separated by a diagonal or curved boundary, a cut flow sacrifices efficiency to remain simple.

The Classifier: What Changes

An ML-based classifier — whether a boosted decision tree, a neural network, or another architecture — learns a single output score from many input observables simultaneously. Instead of cutting on each variable independently, you cut once on the score. This lets the model exploit correlations between observables that a sequential cut flow cannot capture.

In particle physics language: the classifier approximates the optimal test statistic for separating your signal hypothesis from background. If your simulation is a good description of data, this is the most powerful selection you can make, in the Neyman-Pearson sense. You get a better ROC curve — higher signal efficiency at the same background rejection, or equivalently, higher purity at the same efficiency.

That gain is real, but it comes with responsibilities.

What the Switch Actually Costs

Interpretability

A gradient-boosted tree with a few hundred leaves is not as readable as a cut table. A deep network is less readable still. You can use feature-importance tools and SHAP values to understand what the model learned, but the explanation is always post-hoc. Plan for this when writing your analysis documentation and answering internal review questions.

Simulation dependence

A classifier trained on simulation can learn features that are well-modelled and features that are not. A poorly modelled input variable can introduce a sculpting effect on your background estimate that is invisible until you look for it explicitly. Validating a classifier in data-driven control regions is not optional — it is the price of admission.

Overtraining

The ML equivalent of overfitting a chi-square to statistical fluctuations. If your training sample is small relative to the model capacity, the classifier will perform optimistically on training data and disappointingly on test data and real events. Cross-validation and dedicated test sets are your standard tools here. The full HEP ML course covers this workflow in detail if you want a structured treatment.

A Practical Decision Framework

Use a cut flow when:

• Your signal region is well-defined by a small number of physically motivated variables
• You need the selection to be fully reproducible without running code
• Your simulation modelling of candidate input variables has known, large uncertainties
• You are in an early analysis stage and want to iterate quickly

Consider a classifier when:

• Your signal is genuinely separated from background only in high-dimensional feature space
• You have a large, well-validated simulation sample with good data agreement
• You have control regions to validate the classifier output in data
• The analysis is mature enough to carry the additional documentation and systematic evaluation burden

The two approaches are also not mutually exclusive. A common pattern is to apply a loose cut flow to define a clean region, then train a classifier within that region. This limits the phase space the classifier needs to learn and reduces exposure to poorly modelled tails.

Getting Started

If you are new to ML-based event selection, starting with a boosted decision tree on a handful of well-understood variables is a reasonable first step. It is interpretable enough to debug, powerful enough to see a genuine improvement, and close enough to a cut flow that your collaboration will follow what you did. From there, the path to more expressive models is incremental.

The complete course covers the full pipeline from feature engineering through classifier validation in a particle physics context — practical, not abstract.

Takeaway: A classifier is not a replacement for understanding your signal; it is a way to act on that understanding more efficiently once you do.

References

Radovic, A., Williams, M., Rousseau, D., Kagan, M., Bonacorsi, D., Himmel, A., Aurisano, A., Terao, K., & Wongjirad, T. (2018). Machine learning at the energy and intensity frontiers of particle physics. Nature, 560, 41-48.