Inspect per-class performance
Model evaluation
Scikit-Learn · Function
classification_report
Builds a per-class classification-metrics report from true and predicted labels. It returns formatted text by default or structured data when output_dict=True.
How it works
classification_report compares ground-truth labels with classifier predictions and summarizes precision, recall, F1 score, and support for every included class. It can format the result as text or return a dictionary for programmatic inspection.
Syntax
sklearn.metrics.classification_report(y_true, y_pred, *, labels=None, target_names=None, sample_weight=None, digits=2, output_dict=False, zero_division='warn')Parameters
| Name | Type | Description |
|---|---|---|
y_true | 1d array-like, or label indicator array / sparse matrix | Ground-truth target values; supports one-dimensional labels or label-indicator data. |
y_pred | 1d array-like, or label indicator array / sparse matrix | Estimated targets produced by a classifier. |
labels | array-like, default=None | Labels to include in the report. |
target_names | list of str, default=None | Display names corresponding to labels in the same order. |
sample_weight | array-like of shape (n_samples,), default=None | Per-sample weights. |
digits | int, default=2 | Decimal precision for formatted output. |
output_dict | bool, default=False | Return a dictionary instead of formatted text. |
zero_division | warn, 0.0, 1.0, or np.nan, default='warn' | Controls handling of zero-division cases. |
Returns
str or dict: a text summary, or a dictionary when output_dict=True.
Key points
- The default result is a human-readable report of the principal classification metrics.
- Each class receives precision, recall, F1 score, and support information.
- Use labels to restrict which class labels appear in the report.
- Use target_names to replace raw labels with display names, preserving label order.
- digits changes the number of decimal places used when formatting numeric text output.
- The report includes a macro average alongside other applicable aggregate averages.
Common mistakes
- Sparse target matrices are supported only for multilabel targets.
- Set output_dict=True when downstream code needs structured values rather than formatted text.
- The sample average is available only for multilabel classification.
Example: Print a classification report
from sklearn.metrics import classification_report
y_true = [0, 1, 2, 2, 2]
y_pred = [0, 0, 2, 2, 1]
print(classification_report(y_true, y_pred, target_names=["class 0", "class 1", "class 2"]))Prints a readable per-class precision, recall, F1-score, and support report.
When to use it
Compare macro and weighted averages
Imbalanced classification
Consume structured metric data
Evaluation pipelines
Review sample-level averages
Multilabel classification
Interview questions
QWhat does classification_report compute?
It accepts true labels and predicted labels, then reports precision, recall, F1 score, and support for each class. By default it returns a formatted text summary.
Follow-up: How can code consume the metrics without parsing text? Set output_dict=True, which returns a dictionary.
QHow do the reported averages differ?
The macro average is the unweighted mean across labels, while the weighted average weights each label by its support. A sample average applies only to multilabel classification.
Follow-up: When can a micro average appear? It is shown for multilabel data or multiclass data restricted to a class subset, where it is not equivalent to ordinary accuracy.
QHow do labels, target_names, and digits affect the output?
labels selects the label indices included in the report, and target_names supplies display names in the same order. digits controls formatted floating-point precision for text output.
Follow-up: What should be checked when custom names are supplied? Ensure target_names corresponds positionally to the selected labels.
Related
precision_recall_fscore_supportmultilabel_confusion_matrixaccuracy_scoreConfusionMatrixDisplay