BDT models for low pT electron seeds

[mverzett]

This PR introduce the BDT models used in the GSF Seeding in the upcoming PR for low-pT electron reconstruction.
Two models are provided:

1. A model that is aware of the different expected pT and eta spectrum of electrons from B decays and of their different displacement.
2. A model where the background sample is reweighted to remove any discriminating power of pt and eta. Displacement is not provided.

The performance of these two models is compared to the current seeding approach (blue star) and an hypothetical seeding approach same selection as the current but with no minimal pT threshold. The triangle marker show the performance of the OR of the two trained models applying a cut that induces the same mistag rate (respectively 1, 3, and 10 times the mistag rate of the current seeding).

The performance of the same models have been tested applying a posteriori a 0.5 GeV pT cut on the track, showing the same performance in the interested phase space. It is therefore safe to apply such a cut before computing the BDT outputs to save time.

@bainbrid is also interested in following this thread.

[cmsbuild]

A new Pull Request was created by @mverzett (Mauro Verzetti) for branch master.

@cmsbuild, @smuzaffar, @gudrutis, @mrodozov can you please review it and eventually sign? Thanks.
You can sign-off by replying to this message having '+1' in the first line of your reply.
You can reject by replying to this message having '-1' in the first line of your reply.

external issue cms-sw/cmsdist#4563

[slava77]

@mverzett
do we really need this fancy naming "improvedfullseeding_formatted" and perhaps even the date are not really functional and appear to reflect only the history of derivation.
I can imagine this way the next version will be "evenmoreimproved" 😄

[mverzett]

Hi Slava, the name can be changed, please let me know. They reflected some studies we made with different feature sets to understand how much we could gain by adding more and more information. The date represent the internal ntuples used to run the training for internal bookkeeping.

[slava77]

I think that MC sample/campaign tag is more useful in the name than just a simple data which has a meaning only to you.

Do you expect to use both files in reco or are they integrated now "for upcoming studies", not for production?

[slava77]

@mverzett
what is the status of the pull request for CMSSW to use these training files?

[bainbrid]

Imminent. Today or Monday ...

…

On 7 Dec 2018, at 16:47, Slava Krutelyov ***@***.***> wrote: @mverzett <https://github.com/mverzett> what is the status of the pull request for CMSSW to use these training files? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#10 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABEfkpxQPjKIBvyPa6Xokf8jyw1TJAztks5u2puagaJpZM4ZIr8->.

[slava77]

Imminent. Today or Monday ...
…

OK, it sounds like we better wait for that to show up before merging this PR

[slava77]

@mverzett
please let me know if you can rename the files to be a bit more functional

[mverzett]

@slava77 I just realised I made the PR from the wrong repo, I will close and re-open with the names changed