Feature sets: mappings and purpose of FNCv3

In the features.json file that comes in the V4 data we have the following feature set sizes:

  • small: 38
  • medium: 472
  • v2_equivalent_features: 304
  • v3_equivalent_features: 1050
  • fncv3_features: 420

Does anyone know the mapping between the 304 V2 features and the original 310 legacy features?

A further question: what is the purpose of the FNCv3 features? I’ve looked for documentation on this but cannot find it.

Maybe this can help for the first part of your question

2 Likes

Thanks, great research there

Adding a link to the original FNC announcement post for those who are wondering what FNC is about: Feature Neutral Correlation Added to the Tournament Site

Original forum post of Feature Exposure by jrb: Model Diagnostics: Feature Exposure