Why does histogram-based split evaluation still require a pass over observations even though candidate scanning is…
Classic MLLevel: IntermediateLightGBM · Features
Why does histogram-based split evaluation still require a pass over observations even though candidate scanning is O(number of bins)?
Hints
- Separate building and scanning the histogram.
- The compact summary has to come from data first.
Answer
The histogram must first aggregate row statistics. Only scanning the completed histogram scales with bin count; confusing these phases understates total training complexity.
Source and licence: LightGBM · Features · MIT · answer: Quantoros editorial