Wing Sheung Chan

74 Event selection and classification Table 4.3.: Purities of the fakes-enriched regions. Fakes-enriched region Purity [%] eτ channel µτ channel FRW 95 94 FRQ 90 89 FRZ 99 99 FRT 96 96 The purity of an FR in the targeted process is defined as the expected relative contribution of the targeted process to the total amount of fakes in the region. The estimated purities of the FRs are shown in Table 4.3, which are generally high. Table 4.4 summarises the event selection criteria for CRZ τ τ , CRZ `` and the FRs. The purpose and usage of these regions will be discussed in more details in the following chapters. Additionally, several regions are defined particularly for validating the modelling of fakes. They include a validation region, VRSS, that has the same event selection criteria as the SR except that SS instead of OS charged ` – τ had - vis pairs are required. Three special FRs are also defined, they are defined in the same way as FRW, FRZ and FRT, but again requires SS instead of OS charged ` – τ had - vis pairs. They are referred to as the SS FRW, SS FRZ and SS FRT respectively, and are collectively referred to as the SS FRs together with FRQ. These regions are all dominated by fakes. 4.2. Neural network classifiers Neural network (NN) classifier s † are used to optimise the discrimination between the signal and different background events. The utilisation of NN classifiers is at the heart of the analysis. Not only do they allow us to identify the most signal-like events, but they also provide separation between different background processes, which is valuable for improving the precision and accuracy of the background modelling. Multiple binary classifiers are trained for each channel to discriminate signal events from the three major backgrounds: Z → τ τ , W ( → `ν ) +jets and Z → `` events. They are labelled NN Z ττ , NN Wjets and NN Z `` respectively. The output scores of these classifiers are then combined into one final discriminant, the distribution of which is used in binned maximum-likelihood fits in the SR to extract evidence of potential signals or to set upper limits on the LFV branching fractions. † See Appendix A for a brief introduction to neural network classification.