Why should featurefractionbynode not be assumed to provide the same training speedup as featurefraction?
Classic MLLevel: IntermediateLightGBM · Parameters
Why should feature_fraction_bynode not be assumed to provide the same training speedup as feature_fraction?
Hints
- When are histograms built relative to node sampling?
- Fewer eligible splits do not imply proportionally less work.
Answer
Node-level subsampling can regularize split selection without avoiding all tree-level histogram work. The documentation distinguishes their speed effects; measure throughput rather than multiplying feature fractions into an assumed speedup.
Source and licence: LightGBM · Parameters · MIT · answer: Quantoros editorial