colmmm/thetwobbtagger

CERN Summer Student Project 2019

2B^2 Tagger

Summary

A new approach to B-taging, which differs from convential tagging techniques such as the OS-tagger/SS-tagger (OppositeSide/SameSide) by starting off with just two track body (TB) vertex's. This approach also implements tools used in the analysis of Semileptonic decays of B Baryons, which includes the Isolation MVA technique which can be used to find as much of the decay chain of the B as possible. Another technique leveraged from the previous described analysis is the Line-of-Flight (LoF) technique which allowes the estimation of COM quantites of the tracks associated in the decay chain. Then using all the information available to us about the two track vertex and the extra tracks (ETs) added(including information obtained from the LoF technique), a tag (B or BBar) is predicted with a corresponding mistag probability.

thetwobbtagger's Issues

Try LSTM approach for tagging

This would be for the tagging (third) stage, where instead of treating the TBs independently when we feed them into the classifier, we instead group them by their event, and feed them into an LSTM recurrent neural net. So in this approach, the LSTM can handle events with varying number of TBs

Implement ML for the PID classifier within the LOF calculation

Currently, we take the max probability outputed by the NNs on the particle identification (PID), with additional logic that if muon NN being above some threshold we take particle to be a muon. I did try a ML approach using the probs as features but had worse performance

Produce graphs showing LOF features on much more data

Adjust pipeline so it can handle big data flow

optimisation of different stages threshold

First Stage (TBs) and Second Stage (ETs) have thresholds relating to how strict theyre on what TBs/ETs to keep, need to optimise these

Dealing with TBs with no ETs

Currently, as a quick fix I just looked at the TBs which didnt have ETs and were causing problems, as there were only 3 of them I manually removed them by their id but will need to generalise this