Implement and Optimize CatBoost Classifiers

CatBoost Classifier Expert guides the implementation, optimization, and deployment of CatBoost classifiers. It covers the training workflow as well as the decisions that affect categorical features, validation, prediction, and model interpretation.

The skill explains native categorical handling through ordered target statistics and random permutations, symmetric oblivious decision trees, ordered boosting, GPU acceleration, and built-in regularization. These principles are connected to Python implementation patterns rather than left as a separate algorithm overview.

Basic training includes a train/test split, explicit categorical feature names or indices, an AUC evaluation metric, and early stopping against an evaluation set. Advanced Pool configuration adds feature names and optional sample weights, keeping the feature specification attached to the data used for training and evaluation.

Use this skill when developing a CatBoost classification workflow, tuning an existing classifier, or preparing a model for production inference. High-cardinality categories receive specific attention through CTR configuration, combination complexity, and per-feature quantization.

Training guidance emphasizes explicit categorical feature declarations, cross-validation for parameter selection, multiple evaluation metrics, and early stopping. It also addresses imbalanced datasets through weighting and checks that categorical features remain consistent between training and inference.

Large-data settings include GPU task selection, device configuration, CPU thread limits, RAM limits, and model-size regularization. These are configuration patterns to profile and adjust for the dataset, rather than benchmark results or a promised training speed.

Inputs include training and evaluation features, target labels, categorical feature identifiers, optional sample weights, and parameter search ranges. Outputs include fitted classifiers, probability predictions, evaluation results, parameter search results, and exported models.

GridSearchCV examples search iterations, learning rate, depth, leaf regularization, numerical borders, and bagging temperature with ROC AUC scoring. Optuna examples define a trial objective, train with early stopping, and maximize the AUC of predicted probabilities. Custom metric examples include F1, AUC, precision, and recall.

Interpretation uses feature importance tables and SHAP values. The examples sort importance values by feature and use a tree explainer with a summary plot to examine contributions across test observations.

The workflow uses pandas for tabular inputs and scikit-learn for splitting, metrics, and grid search. Optuna provides Bayesian optimization examples, and SHAP provides detailed interpretation.

Deployment examples save native CBM, JSON, and ONNX files and reload a model for prediction. A predictor wrapper uses Pool for batch inference and converts a single feature dictionary into a DataFrame before calling the same batch path.

loaded_model = CatBoostClassifier()
loaded_model.load_model('catboost_model.cbm')