How to create a classification report with scikit-learn

Class-level metrics show where a classifier performs well and where it misses specific labels. In scikit-learn, classification_report() turns held-out labels and predictions into precision, recall, F1-score, and support rows that are easier to inspect than a single accuracy value.

The report takes the true labels and predicted labels from the same evaluation split. Passing target_names keeps rows readable when labels are encoded as integers, and zero_division=0 makes unattended checks deterministic if a class has no predicted samples.

Use the text output for notebook review, pull output_dict=True when a pipeline needs numeric metric values, and read the aggregate rows alongside per-class support. A high weighted average can hide a weak minority-class row, so the support counts matter whenever classes are imbalanced.

1. Save a smoke script that fits a classifier and keeps the held-out labels with the predictions.

   classification_report_demo.py

   from sklearn.datasets import load_wine
   from sklearn.metrics import classification_report
   from sklearn.model_selection import train_test_split
   from sklearn.tree import DecisionTreeClassifier
    
   wine = load_wine()
   X_train, X_test, y_train, y_test = train_test_split(
       wine.data,
       wine.target,
       test_size=0.30,
       stratify=wine.target,
       random_state=7,
   )
    
   model = DecisionTreeClassifier(max_depth=3, random_state=7)
   model.fit(X_train, y_train)
    
   y_pred = model.predict(X_test)
    
   print(
       classification_report(
           y_test,
           y_pred,
           target_names=wine.target_names,
           digits=3,
           zero_division=0,
       )
   )
    
   metrics = classification_report(
       y_test,
       y_pred,
       target_names=wine.target_names,
       output_dict=True,
       zero_division=0,
   )
    
   print(f"macro avg f1-score: {metrics['macro avg']['f1-score']:.3f}")

   Replace the load_wine() split with the project's existing test labels and predictions when the classifier is already trained.
   Related: How to split data into train and test sets with scikit-learn

2. Run the script.

   $ python3 classification_report_demo.py
                 precision    recall  f1-score   support
   
        class_0      1.000     0.833     0.909        18
        class_1      0.870     0.952     0.909        21
        class_2      0.938     1.000     0.968        15
   
       accuracy                          0.926        54
      macro avg      0.936     0.929     0.929        54
   weighted avg      0.932     0.926     0.925        54
   
   macro avg f1-score: 0.929

3. Use the dictionary report when another program needs a metric value instead of formatted text.

   metrics = classification_report(
       y_test,
       y_pred,
       target_names=wine.target_names,
       output_dict=True,
       zero_division=0,
   )
    
   print(f"macro avg f1-score: {metrics['macro avg']['f1-score']:.3f}")

   digits affects only the formatted text report. The dictionary returned by output_dict=True keeps unrounded numeric values.

4. Confirm that every expected class has a row in the report.

   The support values show how many held-out samples belong to each class. In the sample output, 18 + 21 + 15 matches the 54 samples shown on the accuracy and aggregate rows.