On the other hand, recall goes down the stronger the filter is, which is natural as a strong filter is likely to remove informative examples about less frequently used endpoints. It has to be noted, though, that recall remains largely unaffected for considerable strong noise filters, indicating a relatively broad range in which it can be applied without negative impacts. Overall, the application of a noise filter improves the overall F1 score from around 0.46 up to 0.64. When applying our method to an unknown dataset though, the impact may differ. Possibly, an adequate noise filter level might be learned over time from experiments.