Why can using fewer histogram bins reduce memory as well as split-search time?
Classic MLLevel: IntermediateLightGBM · Features
Why can using fewer histogram bins reduce memory as well as split-search time?
Hints
- Raw floating-point values can be replaced by bin indices.
- Representation size and resolution trade off.
Answer
Quantized feature values can be stored as compact bin IDs, sometimes in smaller integer types. This saves storage but reduces threshold resolution, so benchmark quality near important decision boundaries.
Source and licence: LightGBM · Features · MIT · answer: Quantoros editorial