Developing Jet Vertex Tagger in larger pT range using neural network

The Jet Vertex Tagger (JVT) is designed for separating Hard-Scattering Jets from Pileup Jets, using machine learning techniques. Originally, a KNN-based algorithm was trained for the classification of the jet with in a pT range (20-50 GeV). The objective of this study is to extend it to a larger pT range (20-120 GeV). We developed a Track-Based-Dynamic-Neural-Network (TBDNN) to classify jets using their track information. This algorithm use the momentum of jets and tracks as input instead of the observables constructed from them, avoiding some potential information loss. The layers in the network are dynamically reweighted for each single input jet so that the network could automatically adjust itself for jets with different track number. When applied to simulated data, TBDNN has shown a considerable advantage in high pT range (50-120 GeV) compared with K-Nearest-Neighbors (KNN). The new algorithm may benefit the MET by improving the rejection of high pT pileup jets.