Learn AI Create the Future — with Pranay
Main site

Error database

ValueError: Classification metrics can't handle a mix of binary and continuous targets

You passed probabilities or regression outputs into a metric that expects class labels. Threshold or argmax the predictions first — or switch to a metric that wants scores.

The message you saw
ValueError: Classification metrics can't handle a mix of binary and continuous targets

By Updated

The error

Output
ValueError: Classification metrics can't handle a mix of binary and continuous targets

Variants swap the two type words: multiclass and continuous, binary and continuous-multioutput.

What it means

A metric like accuracy_score compares two lists of class labels: truth versus prediction, both discrete. One of the lists you passed contains continuous numbers — probabilities like 0.73, or regression outputs like 12.4. The metric names both types it saw; the second one is usually the intruder.

Why it happens

Three common routes:

  • You called model.predict_proba(X) and fed the probabilities to accuracy_score. Probabilities are not labels yet.
  • You trained a regressor on a yes/no problem, so even .predict returns continuous values.
  • With neural networks, the model outputs one probability (or several logits) per class and you skipped the conversion to a label.

How to fix it

1. Use .predict when the metric wants labels.

python
from sklearn.metrics import accuracy_score
preds = model.predict(X_test)              # labels, not probabilities
print(accuracy_score(y_test, preds))

2. Converting probabilities yourself? Threshold or argmax.

python
proba = model.predict_proba(X_test)
preds = (proba[:, 1] >= 0.5).astype(int)   # binary: threshold column 1
python
import numpy as np
preds = np.argmax(proba, axis=1)           # multiclass: highest-probability class

3. Some metrics want the scores — do not convert for those. roc_auc_score and log_loss measure ranking and calibration; give them probabilities:

python
from sklearn.metrics import roc_auc_score
print(roc_auc_score(y_test, proba[:, 1]))

The rule: accuracy, precision, recall, F1, confusion matrix take labels; ROC-AUC and log loss take scores.

4. If a regressor is answering a classification question, change the model. Use a classifier and this whole ambiguity disappears — see the related error below for that trap.

How to prevent it

Print the first five values of whatever you pass to a metric. Labels look like [0, 1, 1, 0, 2]; scores look like [0.73, 0.11, ...]. Match each metric to its input type on purpose, and keep the conversion (threshold or argmax) as one explicit, visible line.