Beyond cuts in small signal scenarios

Abstract We investigate enhancing the sensitivity of new physics searches at LHC by machine learning in case background dominance and a high degree overlap between observables for signal background. use two different models, XGBoost deep neural network, to exploit correlations compare this approach traditional cut-and-count method. consider methods analyze models’ output, finding that template fit generally performs better than simple cut. By means Shapley decomposition, we gain additional insight into relationship event kinematics model output. supersymmetric scenario with metastable sneutrino as concrete example, but methodology can be applied much wider class models.