Mass-decorrelated Xbb Tagger using Adversarial Neural Network

One key task performed by the ATLAS experiment at the LHC is the Xbb tagging, which refers to the identification of Higgs bosons decaying into bottom quark pairs ($H$ → $b$$\bar{b}$. Deep neural networks (DNN) have been adopted to develop Xbb taggers. While DNN-based taggers are generally performant in signal vs. background classification, they tend to have high jet mass correlation, which is undesired as it sculpts the jet mass distribution of the background to resemble that of the signal. One technique to minimize jet mass correlation is the training of an adversarial neural network (ANN). In this study we demonstrate the application of ANN training to develop two mass-decorrelated Xbb taggers (Hbb vs. Dijet and Hbb vs. Top). The performance is evaluated with metrics for classification power and jet mass correlation