CatBoost (Categorical Boosting) is a gradient boosting library developed by Yandex that provides native, leak-free handling of categorical features without any manual encoding. In most real-world datasets, a significant proportion of features are categorical — product categories, customer segments, geographic regions, contract types — and traditional approaches (one-hot encoding, label encoding, target encoding) each introduce their own problems. One-hot encoding creates high-dimensional sparse matrices for high-cardinality categoricals. Label encoding imposes false ordinal structure. Target encoding leaks the target into the features if not handled carefully with out-of-fold encoding.
CatBoost addresses all these problems simultaneously with ordered target encoding (also called ordered boosting): it applies target encoding to categorical features but computes each example's encoding using only the training examples that came before it in a random ordering, preventing any target information from leaking from each example into its own encoding. This leak-free categorical encoding combined with symmetric tree growth and strong default regularisation make CatBoost particularly competitive on tabular datasets with many categorical features. Understanding when to use CatBoost versus LightGBM versus XGBoost requires understanding how each handles categorical features and what constitutes the dataset's dominant challenge.