This document discusses the use of machine learning and scikit-learn in particle physics experiments like those done at CERN. It describes how machine learning techniques were very successful in the Kaggle Higgs Boson challenge at classifying particle interactions, outperforming approaches used by physicists. However, scikit-learn has not been widely adopted in particle physics due to technical limitations and social/cultural reasons within the community. The document argues that greater collaboration between machine learning experts and physicists could help overcome these issues and make scikit-learn a more important tool for data analysis in high energy physics.

1. Scikit-Learn in particle physics
Gilles Louppe
CERN, Switzerland
November 18, 2014
1 / 13

2. High Energy Physics (HEP)
c CERN
2 / 13

3. High Energy Physics (HEP)
c CERN
Study the nature of the
constituents of matter
2 / 13

4. High Energy Physics (HEP)
c CERN
Study the nature of the
constituents of matter
2 / 13

5. Particle detector 101
c ATLAS CERN
3 / 13

6. Particle detector 101
c ATLAS CERN
3 / 13

7. Data analysis tasks in detectors
1 Track

8. nding
Reconstruction of particle trajectories from hits in detectors
2 Budgeted classi

9. cation
Real-time classi

10. cation of events in triggers
3 Classi

11. cation of signal / background events
Oine statistical analysis for discovery of new particles
4 / 13

12. The Kaggle Higgs Boson challenge (in HEP terms)
Data comes as a

13. nite set
D = f(xi ; yi ;wi )ji = 0; : : : ;N 1g;
where xi 2 Rd , yi 2 fsignal; backgroundg and wi 2 R+.
The goal is to

14. nd a region G = fxjg(x) = signalg Rd ,
de

15. ned from a binary function g, for which the
background-only hypothesis can be rejected at a strong
signi

16. cance level (p = 2:87 107, i.e., 5 sigma).
Empirically, this is approximately equivalent to

17. nding g from
D so as to maximize AMS ps
b
, where
s =
P
fi jyi =signal;g(xi)=signalg wi
b =
P
fi jyi =background;g(xi)=signalg wi
5 / 13

18. The Kaggle Higgs Boson challenge (in ML terms)
Find a binary classi

19. er
g : Rd7! fsignal; backgroundg
maximizing the objective function
AMS
s
p
b
;
where
s is the weighted number of true positives
b is the weighted number of false positives.
6 / 13

20. Winning methods
Ensembles of neural networks (1st and 3rd) ;
Ensembles of regularized greedy forests (2nd) ;
Boosting with regularization (XGBoost package).
Most contestants dit not optimize AMS directly ;
But chosed the prediction cut-o maximizing AMS in CV.
7 / 13

21. Lessons learned (for machine learning)
AMS is highly unstable, hence the need for
Rigorous and stable cross-validation to avoid over

22. tting.
Ensembles to reduce variance ;
Regularized base models.
Support of samples weights wi in classi

23. cation models was
key for this challenge.
Feature engineering hardly helped. (Because features already
incorporated physics knowledge.)
8 / 13

24. Lessons learned (for physicists)
Domain knowledge hardly helped.
Standard machine learning techniques, run on a single laptop,
beat benchmarks without much eorts.
Physicists started to realize that collaborations with machine
learning experts is likely to be bene

25. cial.
I worked on the ATLAS experiment for over a decade [...] It is rather
depressing to see how badly I scored.
The

26. nal results seem to reinforce the idea that the machine learning
experience is vastly more important in a similar contest than the
knowledge of particle physics. I think that many people underestimate the
computers.
It is probably the reason why ML experts and physicists should work
together for

27. nding the Higgs.
9 / 13

28. Scienti

29. c software in HEP
ROOT and TMVA are standard data analysis tools in HEP.
Surprisingly, this HEP software ecosystem proved to be rather
limited and easily outperformed (at least in the context of the
Kaggle challenge).
10 / 13

30. Scikit-Learn in Particle Physics ?
The main technical blocker for the larger adoption of
Scikit-Learn in HEP remains the full support of sample
weights throughout all essential modules.
Since 0.16, weights are supported in all ensembles and in most metrics.
Next step is to add support in grid search.
In parallel, domain-speci

31. c packages are getting traction
ROOTpy, for bridging the gap between ROOT data format and NumPy ;
lhcb trigger ml, implementing ML algorithms for HEP (mostly Boosting
variants), on top of scikit-learn.
11 / 13

32. Major blocker : social reasons ?
The adoption of external solutions (e.g., the scienti

33. c Python
stack or Scikit-Learn) appears to be slow and dicult in the HEP
community because of
No ground-breaking added-value ;
The learning curve of new tools ;
Lack of understanding of non-HEP methods ;
Isolation from the community ;
Genuine ignorance.
12 / 13

34. Conclusions
Scikit-Learn has the potential to become an important tool in
HEP. But we are not there yet [WIP].
Overall, both for data analysis and software aspects, this calls
for a larger collaboration between data sciences and HEP.
The process of attempting as a physicist to compete against ML
experts has given us a new respect for a

35. eld that (through
ignorance) none of us held in as high esteem as we do now.
13 / 13

36. c xkcd
Questions ?
14 / 13