CatBoost

CatBoost (Yandex) is arguably the easiest boosting library to use because it handles Categorical Features automatically and perfectly without tuning.

When to Use

• Tabular Data with Categorical Features: Automatic target encoding and categorical combination without manual one-hot encoding.
• Low-Latency Inference: Fast model scoring on CPU and GPU with minimal memory overhead.
• Ranking & Classification Tasks: E-commerce recommendation, fraud detection, and click-through-rate (CTR) prediction.
• Robustness Against Overfitting: Symmetric (oblivious) decision trees providing stable predictions on small-to-medium datasets.

Quick Start

from catboost import CatBoostClassifier, Pool
from sklearn.model_selection import train_test_split
import pandas as pd

# Native categorical feature support
df = pd.DataFrame({
    'city': ['London', 'Paris', 'Tokyo', 'London', 'Paris'],
    'age': [25, 40, 30, 22, 55],
    'target': [1, 0, 1, 1, 0]
})

X = df[['city', 'age']]
y = df['target']
cat_features = ['city']

model = CatBoostClassifier(iterations=100, learning_rate=0.1, verbose=20)
model.fit(X, y, cat_features=cat_features)
preds = model.predict(X)

Core Concepts

Native Categorical Feature Handling

Training CatBoostClassifier directly on raw categorical string columns:

from catboost import CatBoostClassifier, Pool
from sklearn.model_selection import train_test_split
import pandas as pd

# Load dataset with categorical features
data = pd.DataFrame({
    'city': ['London', 'Paris', 'Tokyo', 'London', 'Tokyo', 'Berlin'],
    'device': ['mobile', 'desktop', 'mobile', 'tablet', 'desktop', 'mobile'],
    'age': [24, 45, 32, 19, 58, 29],
    'churned': [0, 1, 0, 0, 1, 0]
})

X = data[['city', 'device', 'age']]
y = data['churned']
cat_features = ['city', 'device']

X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.33, random_state=42)

train_pool = Pool(X_train, y_train, cat_features=cat_features)
val_pool = Pool(X_val, y_val, cat_features=cat_features)

model = CatBoostClassifier(
    iterations=500,
    learning_rate=0.05,
    depth=6,
    eval_metric='AUC',
    early_stopping_rounds=30,
    random_seed=42,
    verbose=50
)

model.fit(train_pool, eval_set=val_pool)

Feature Importance & SHAP Values

Interpreting model decisions with Tree SHAP:

import shap

# Compute SHAP values directly from CatBoost
explainer = shap.TreeExplainer(model)
shap_values = explainer(val_pool)

# Summarize feature impact
feature_importance = model.get_feature_importance(val_pool, type='PredictionValuesChange')
for name, score in zip(X.columns, feature_importance):
    print(f"Feature: {name:10s} Importance: {score:.4f}")

Exporting Model to ONNX for High-Speed Serving

Serializing trained models for cross-platform C++/Go/Rust inference:

# Save model in ONNX format
model.save_model("catboost_model.onnx", format="onnx")

# Or native binary format
model.save_model("catboost_model.cbm")

Common Patterns

Early Stopping with Validation Pool

Problem: Overfitting on tabular training data during deep gradient tree boosting.

Solution: Pass validation pool with early stopping rounds:

train_pool = Pool(X_train, y_train, cat_features=cat_features)
eval_pool = Pool(X_val, y_val, cat_features=cat_features)

model = CatBoostClassifier(
    iterations=1000,
    early_stopping_rounds=50,
    eval_metric='AUC',
    random_seed=42
)
model.fit(train_pool, eval_set=eval_pool, verbose=100)
print(f"Best iteration: {model.get_best_iteration()}")

Best Practices

Do:

• Pass categorical columns directly to cat_features rather than manual one-hot encoding.
• Use early_stopping_rounds with a dedicated validation set to prevent overfitting.
• Train on GPU (task_type="GPU") for datasets exceeding 1 million rows for up to 10x-20x speedup.
• Monitor evaluation metrics using plot=True in Jupyter or export training logs to TensorBoard.

Don't:

• Label encode high-cardinality categoricals manually; CatBoost's target encoding is statistically superior.
• Use too large depth (> 10) unless explicitly regularized; CatBoost defaults (depth 6) are optimal.
• Evaluate final performance on the validation set used for early stopping; evaluate on a held-out test set.

Troubleshooting

References

• CatBoost Documentation