CatBoost报错：无法计算需对数概率的绝对值相关指标

Let's break down what's causing this error and how to fix it quickly.

The Root Cause #

Your error stems from passing use_best_model=True (via cat.get_params()) to the cv() function. This parameter is designed for single-model training—where you have a fixed train/validation split to pick the best iteration—but it doesn't make sense in cross-validation. CatBoost handles fold-specific validation internally, so keeping this parameter triggers a mismatch that breaks metric calculation when using eval_metric='Accuracy'.

Solution 1: Remove use_best_model from CV Parameters #

Extract your model parameters, remove the incompatible use_best_model key, then pass the cleaned params to cv():

from catboost import Pool, CatBoostClassifier, cv
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
import numpy as np

# Split data
train = data[:split]
test = data[split:]
# Get variables for a model
x = train.drop(["Survived"], axis=1)
y = train["Survived"]
# Do train data splitting
X_train, X_test, y_train, y_test = train_test_split(x,y, test_size=0.2, random_state=42)
cat_features = np.where(x.dtypes != float)[0]

# Initialize model with use_best_model (fine for fit(), not for cv())
cat = CatBoostClassifier(one_hot_max_size=7, iterations=21, random_seed=42, use_best_model=True, eval_metric='Accuracy')
cat.fit(X_train, y_train, cat_features = cat_features, eval_set=(X_test, y_test))
pred = cat.predict(X_test)

# Prepare pool and cleaned params for CV
pool = Pool(X_train, y_train, cat_features=cat_features)
cv_params = cat.get_params()
cv_params.pop('use_best_model')  # Remove the incompatible parameter

cv_scores = cv(pool, cv_params, fold_count=10, plot=True)
print('CV score: {:.5f}'.format(cv_scores['test-Accuracy-mean'].values[-1]))
print('The test accuracy is :{:.6f}'.format(accuracy_score(y_test, cat.predict(X_test))))

Solution 2: Move use_best_model to the fit() Call #

Alternatively, only specify use_best_model when calling fit() (since that's where it's actually needed), rather than including it in the model's initial parameters:

from catboost import Pool, CatBoostClassifier, cv
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
import numpy as np

# Split data
train = data[:split]
test = data[split:]
# Get variables for a model
x = train.drop(["Survived"], axis=1)
y = train["Survived"]
# Do train data splitting
X_train, X_test, y_train, y_test = train_test_split(x,y, test_size=0.2, random_state=42)
cat_features = np.where(x.dtypes != float)[0]

# Initialize model WITHOUT use_best_model
cat = CatBoostClassifier(one_hot_max_size=7, iterations=21, random_seed=42, eval_metric='Accuracy')
# Pass use_best_model only to fit()
cat.fit(X_train, y_train, cat_features = cat_features, eval_set=(X_test, y_test), use_best_model=True)
pred = cat.predict(X_test)

pool = Pool(X_train, y_train, cat_features=cat_features)
# Now cat.get_params() won't include use_best_model, so cv() works
cv_scores = cv(pool, cat.get_params(), fold_count=10, plot=True)
print('CV score: {:.5f}'.format(cv_scores['test-Accuracy-mean'].values[-1]))
print('The test accuracy is :{:.6f}'.format(accuracy_score(y_test, cat.predict(X_test))))

Both approaches resolve the conflict by ensuring use_best_model isn't passed to the cross-validation function, where it doesn't belong. The second option is cleaner since it keeps parameters scoped to where they're used.

内容的提问来源于stack exchange，提问作者Stanislav Jirák