Ce Classification

CE Classification

You are working with a classification task (binary or multiclass). CE provides calibrated probability bounds $[p_{low}, p_{high}]$ using Venn-Abers calibration.

Prediction Semantics (Venn-Abers)

The predict and predict_proba methods return calibrated class assignments and probabilities.

• Binary: Rules and probabilities always refer to the positive class (index 1).
• Multiclass:
  • predict(): Returns the class label with the highest calibrated probability.
  • predict_proba(): Returns a full probability distribution (summing to 1).
• Intervals: For any class probability $p$, the calibrated interval $[p_{low}, p_{high}]$ guarantees that $p_{low} \le p \le p_{high}$ (ADR-021 invariant).

# Multiclass: explain the PREDICTED class (default)
# If the model predicts 'Class 2', rules explain 'Why Class 2?'
explanations = explainer.explain_factual(X_test)

Explaining All Classes (multi_labels_enabled)

By default, CE explains the predicted class. To explain all classes (e.g., for contrastive analysis or true multi-label tasks), use multi_labels_enabled=True.

# Returns a MultiClassCalibratedExplanations object
multi_explanations = explainer.explain_factual(
    X_test,
    multi_labels_enabled=True
)

MultiClassCalibratedExplanations API

This object stores a mapping of {class_index: Explanation} for each instance.

# 1. Access instance i (returns a specialized multi-class view for instance i)
inst_exps = multi_explanations[i]

# 2. Access a specific class explanation for instance i
# Returns a standard FactualExplanation or AlternativeExplanation
class_1_exp = multi_explanations[i, 1]

# 3. Plotting
# Plots all class explanations for the instance
multi_explanations[i].plot()

Task-Specific Config

When fitting the explainer, CE automatically detects binary vs multiclass based on y_proper. Use explainer.class_labels to check the detected labels.

explainer = WrapCalibratedExplainer(model)
explainer.fit(x_p, y_p)
explainer.calibrate(x_c, y_c)

print(f"Task: {explainer.task}")             # 'classification'
print(f"Labels: {explainer.class_labels}")   # e.g., ['Setosa', 'Versicolor', 'Virginica']

Contributor & Agent Checklist

1. Binary index — remember that binary classification in CE is always relative to class 1.
2. Multi-label type — use multi_labels_enabled=True (plural labels) to trigger the multiclass explanation logic.
3. Invariant check — verify low <= predict <= high for any class probability being inspected.
4. Labels in titles — when plotting multiclass, ensure class labels (not just indices) are used if available in explainer.class_labels.