We document the performance of the Neural Network based Jet Charge Tagger on a sample of data containing -hadrons with semileptonic decays. The events are triggered by a lepton with transverse momentum larger than 4 GeV and a displaced track with GeV (SVT track). We construct a sample with close to 100% purity by applying a cut on the invariant mass of the lepton and the displaced track and by performing a a background subtraction using the signed impact parameter of the displaced trigger track.
The Jet Charge Tagger is one of the tagging algorithms used in the mixing analyses. Flavor Taggers are used to identify the type of the -hadrons when they are produced. In this case the -hadrons that are studied for mixing fire the CDF trigger and the tracks used for the Jet Charge flavor tag are associated with the second quark produced in the event.
This analysis uses Monte Carlo samples to study the properties of events and optimize the performance of the tagger. For this procedure to be successful the physics and detector performance of the Monte Carlo must reproduce the performance seen in real data. We demonstrate the accuracy of our Monte Carlo by studying a large number of critical distributions to show the high level of data and Monte Carlo agreement. The Monte Carlo sample used is generated with Pythia  and includes all production processes. We are able to extract the -jet properties from the Monte Carlo and use them for the jet selection in the data.
We introduce Neural Network to tag the B decay products with a track probability network. We define several jet variables based on the track probability. The jet variables are combined with a Neural Network as well to give a -jet probability.
The track probability Neural Network (trackNet) has been calibrated on a +SVT Monte Carlo sample and estimates how likely it is that the track is a B decay product. The optimization has been performed with the NeuroBayes  package. The set of input variables chosen for the trackNet is:
The tracks are reconstructed using also the hits in the closest silicon layer to the interaction point, layer 00 (L00). The resolution on and is influenced by the presence of L00 hit on the track. The effect is taken into account by optimizing the trackNet separately for tracks with and without L00 hits. Tracks with a L00 hit have a larger spread of trackNet outputs, due to the higher resolution on .
The jets on the opposite side are reconstructed with a simple fixed cone algorithm ( ) which uses tracks only. The jet probability Neural Network, bJetNet, is optimized on a flag that tells if the jet is a -jet or not. The list of input variables for the bJetNet is:
The jet with the highest probability is selected as the tagging jet. The opposite side flavor is given by the sign of the jet charge computed for the tagging jet as:
The flavour on the trigger side corresponds to the charge of the trigger lepton. The dilution is scaled to take into account mixing on the same side and sequential semileptonic decays. In order to compute we split the tagging jet sample in subsamples (3 bins of jet quality and 11 bins of , where is the jet probability). The jet quality types are: