TopolearnCatBoost often matches or beats tuned alternatives on tabular data with default settings, so it's a strong first model.
CatBoost implements gradient boosting: it builds hundreds or thousands of small decision trees, each one fitted to correct the errors of the ensemble so far. The Python package (pip install catboost) follows scikit-learn conventions:
CatBoostClassifierfor binary and multiclass problems,CatBoostRegressorfor numbers,CatBoostRankerfor rankingmodel.fit(X, y),model.predict(X), and for classifiersmodel.predict_proba(X)Xcan be a pandas DataFrame, a NumPy array or a CatBoostPool
Defaults are good: 1000 trees (iterations), depth 6, and a learning rate picked automatically from the data size and iteration count. Many datasets need little more.
Two practical settings you'll set constantly:
verbose: by default CatBoost prints a line per iteration. Useverbose=0to silence it, orverbose=100to print every 100 iterations.random_seed: training has randomness; fix it for reproducible results.
CatBoost also writes a catboost_info/ folder of training logs by default; set train_dir= to choose the location or allow_writing_files=False to disable it.
Models also work inside scikit-learn tooling such as cross_val_score and Pipeline, though CatBoost's native features (categorical handling, its own cross-validation, early stopping) are usually used directly.
Example
from catboost import CatBoostClassifier, CatBoostRegressor
from sklearn.datasets import load_breast_cancer, load_diabetes
from sklearn.model_selection import train_test_split
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, test_size=0.25, random_state=0)
clf = CatBoostClassifier(iterations=300, verbose=0, random_seed=0,
allow_writing_files=False)
clf.fit(X_tr, y_tr)
print(clf.predict(X_te[:3])) # e.g. [0 1 1] class labels
print(clf.predict_proba(X_te[:1]).round(3)) # [[p_class0 p_class1]]
print(round(clf.score(X_te, y_te), 3)) # accuracy, about 0.98
Xr, yr = load_diabetes(return_X_y=True)
reg = CatBoostRegressor(iterations=300, verbose=0, random_seed=0,
allow_writing_files=False)
reg.fit(Xr[:350], yr[:350])
print(reg.predict(Xr[350:353]).round(1)) # three predicted values
print(reg.get_params()["iterations"], reg.tree_count_) # 300 300Remember
- CatBoostClassifier/Regressor follow the fit/predict/predict_proba API
- Defaults (1000 trees, depth 6, automatic learning rate) are strong
- Use verbose=0 and random_seed for quiet, reproducible runs
- allow_writing_files=False stops the catboost_info log folder
Watch out for
- Leaving the default verbose output on in loops or notebooks, flooding the output with thousands of lines.