Higgs BosonDetector

The Higgs boson turns up in about one collision in a billion, then falls apart right away. This model learns to spot those rare events. Scroll down to see how it works, and try it yourself.

click to fire a collision · move to bend the tracks

scroll to explore_
01

The collisions

250,000 collisions from the ATLAS detector, each already labelled signal (a Higgs) or background (everything else). This is what the model learns from.

collisions to learn from
training set
are signal
85,667 events
physics measurements
17 measured · 13 derived
events to classify
submitted to Kaggle
02

Pick a cutoff

The model scores every event from 0 to 1, where higher means more Higgs-like. You set the cutoff: anything above it gets flagged as signal. Drag it and watch the trade-off. Flag too little and you miss real Higgs events; flag too much and background slips through. The AMS score (short for Approximate Median Significance, the competition's metric) rewards the best balance.

3.58
AMS score
higher is better
35%
Higgs caught
of all real signal
1.1%
false alarms
background flagged

where events land

00.51model score: low = background, high = signal
signal backgroundpeach zone = flagged as signal

AMS at this cutoff

1234cutoff →
cutoff0.037
Drag the line, the curve, or the slider. Measured on a 25% held-out slice; the headline 3.71 is the average across 5 folds.
03

Two models

I trained two kinds of model on the same examples and scored them the same way. The gradient-boosted trees (XGBoost) edged out the neural network (a multilayer perceptron, or MLP) here. That's common for this kind of table-shaped problem.

XGBoostselectedAMS 3.710 ± 0.192
best cutoff 0.062
PyTorch MLPAMS 3.515 ± 0.155
best cutoff 0.021
04

What matters most

These are the measurements the model leaned on most. The estimated mass of the two tau particles stands out by far, which is exactly what you'd expect from the physics.

DER_mass_transverse_met_lep
0.163
DER_mass_MMC
0.123
PRI_tau_pt
0.093
DER_mass_vis
0.070
DER_deltaeta_jet_jet
0.056
DER_pt_ratio_lep_tau
0.047
DER_mass_jet_jet
0.036
DER_met_phi_centrality
0.036
DER_pt_h
0.032
DER_deltar_tau_lep
0.032
DER_prodeta_jet_jet
0.030
PRI_lep_eta
0.025

DER_mass_MMC is the model's estimate of the combined mass of the two tau particles, by far the strongest sign that a collision was a Higgs.