XGBoost

When to Use

• Tabular Data: It usually beats Deep Learning on structured tables.
• Speed: Extremely optimized C++ backend.

Core Concepts

Gradient Boosting

DMatrix

Device Parameter

device="cuda"

Best Practices (2025)

• Use

  device="cuda"

  : GPU training is 10x faster.
• Use Early Stopping: Stop training when validation error rises.
• Pass Polars Dataframes: No need to convert to Pandas/NumPy first.

• Don't use one-hot encoding: Use native categorical support (

  enable_categorical=True

  ).

References

• XGBoost Documentation