CatBoost spends more time constructing Pool than training on a mostly numeric matrix. What source-backed…
Classic MLLevel: IntermediateCatBoost · Pool data contract
CatBoost spends more time constructing Pool than training on a mostly numeric matrix. What source-backed representation experiment is reasonable?
Hints
- Profile data conversion separately from boosting.
- Data representation can dominate short training runs.
Answer
Benchmark contiguous numerical data with float32 or the documented FeaturesData path instead of Python object-heavy structures. Verify precision and categorical schema before changing representation; training tuning alone will not fix ingestion overhead.
Source and licence: CatBoost · Pool data contract · Apache-2.0 · answer: Quantoros editorial