Goal of this competition is to determine the flavour of charged B mesons (mesons that contain
b
or
anti-b
quark).
In simpler words, guess whether particle is B+ or B- (guess, whether the particle is meson or anti-meson).
This is the second challenge in the course of machine learning organized by
[Yandex School of Data Analysis](http://yandexdataschool.com)
at
[Imperial College London](https://www.imperial.ac.uk)
.
Idea of tagging
Mesons are particles that contain a quark and an anti-quark.
[B mesons](https://en.wikipedia.org/wiki/B_meson)
are mesons that contain
b
or
anti-b
quark.
Each B-meson production is accompanied by birth of other particles that leave tracks in the detector. Type and properties of particles and their corresponding tracks are tightly connected with a quark (b or anti-b) forming a meson of interest.
Thus, we can guess the quark in the B-meson by analyzing properties of tracks in the detector.
Bad news: there are many tracks in the detector, some may have no connection to B-meson. Thus, it is a problem with noise.
However, the problem is interesting, because we
[cannot](https://en.wikipedia.org/wiki/Color_confinement)
observe a single quark, but only a hadron (particle consisting of several quarks). Thus, such techniques potentially give us ability to study the quark properties.
No, we don't assume that you know physics.
Data Files
trainingB.root
- training samples with B-meson description
trainingTrack.root
- training samples with other particles in a collision
testB.root
-  test samples with B-meson description
testTrack.root
-  test samples with other particles in a collision
sample_submission.csv
- simple submission, demonstrates expected format
Data Format
trainingB.root
and
testB.root
contain information about the B meson (goal: predict B-meson charge). Column
chargeB
is a target.
trainingTracks.root
and
testTracks.root
contain information about the tracks in the collision. Tracks with the same
ID
corresponds to the same collision and the B-meson corresponding to this collision is in the
trainingB.root
and
testB.root
with the same
ID
.
The file for B-meson looks like:
ID ptB etaB phiB chargeB
0 7.487484 3.598782 2.042920 1
1 1 6.157421 2.812284 4.060636 1
The file for tracks looks like (both tracks connected to the same B-meson with ID=0):
ID ptTrack etaTrack phiTrack track_quality PIDNNe ...
0 0.768184 2.004773 0.183600 1.357150 7.140722e-08 ...
0 0.389962 2.694220 0.216586 2.021095 1.792256e-06 ...
The test samples have the same structure (with the exception of
chargeB
column which is absent in test).
Features Description
B-mesons features:
ID
- collision ID
ptB
- transverse momentum of B-meson
etaB
-
[pseudorapidity](https://en.wikipedia.org/wiki/Pseudorapidity)
of B-meson. It is almost the same as polar angle, see wikipedia
phiB
- azimuth angle of B-meson
chargeB
- the target (is absent in test samples). -1 corresponds to b-quark and 1 correspond to anti-b quark.
For every track registered in the event we calculate several features. Each track has following description:
ID
- collision ID
ptTrack
- transverse momentum
etaTrack
- pseudorapidity
phiTrack
- azimuth angle
track_quality
- quality of track reconstruction
PIDe
- particle identification: probability for this track to come from electron
PIDk
- particle identification: probability for this track to come from kaon
PIDmu
- particle identification: probability for this track to come from muon
PIDpi
- particle identification: probability for this track to come from pion
PIDp
- particle identification: probability for this track to come from proton
PIDghost
- particle identification: probability to be a ghost track (= noise, which was erroneously considered as a track left by particle)
IP
- impact parameter: the distance between the track and vertex from which B-meson appeared
IPerr
- the error of the determination of IP
vertexIP
- impact parameter: the minimum distance between the track and other reconstructed collision vertices.
EOverP
- E/P, where E - energy deposit in the calorimeter system, P - momentum of the particle
Edx
- energy deposit per dx at the beginning of track
chargeTrack
- charge of a track
Discussions
Please discuss the competition in the
[gitter](https://gitter.im/MLatImperial2017/Lobby)
chat of the course.
Submission Format
ID,Prediction
0,0.54
1,0.34
2,0.22
...
The higher predictions should correspond to +1 B-meson charge and lower to -1.
Evaluation
Metric computed on the leaderboard is area under the
[ROC curve](http://arogozhnikov.github.io/RocCurve.html)
(ROC AUC). This metric is discussed in the course.
Validation dataset is split in 2 equal parts: public and private. First is shown on the leaderboard, second is used to compute final standings.
Getting started
The code to start will be published in the
[repository of the course](https://github.com/yandexdataschool/MLatImperial2017)
.
Teams
You can participate in the teams of two!
Only
one
student from physics department can be in the team.
Acknowledgements
We are grateful to the LHCb collaboration at LHC (and personally to Guy Wilkinson) for allowing us to use simulation samples in this competition. Thanks to Denis Derkach for preparing data samples.
Disclaimer
Provided samples were preprocessed for the purposes of the competition. No judgment can be done based on these samples about the properties of LHCb detector.