Skip to content

Add bounds to features, OVL metric reweighting

Patrick Godwin requested to merge ovl_bounds_metrics into master

This PR addresses a few aspects of classifiers as part of p(glitch) tuning investigations, mostly aimed at OVL but some address all classifiers.

All classifiers:

  • Add bounds to all features, not just target features.

OVL:

  • Fix definition of window to be consistent across
  • Modify scaling between metric and rank so that they're more uniform across metrics. This is most notable for the use percentage, where use_percentage = rank.
  • Allow a combined metric where they are averaged (with possibly unequal weights)
  • Reorder OVL list before terminating. Without this, it could lead to ranks not matching up to their metric values for rare veto configurations.

Merge request reports