Skip to content
Quantoros

Why are negative or fractional category IDs risky when passed to LightGBM as categorical features?

Classic MLLevel: IntermediateLightGBM · Advanced topics

Why are negative or fractional category IDs risky when passed to LightGBM as categorical features?

Hints
  • Inspect conversion before the split algorithm sees values.
  • Two inputs may map to one code.

Answer

Negative IDs are treated as missing, and fractional values are truncated toward zero. Distinct intended categories can collapse or disappear; validate nonnegative integer codes and their supported range.

Source and licence: LightGBM · Advanced topics · MIT · answer: Quantoros editorial

Related questions