ham.py

# After review with respect to equations


def hierarchical_precision_recall_fmeasure(
    true_labels, predicted_labels, ancestors, beta=1.0
):
    # Initialize counters for true positives, predicted, and true conditions
    true_positive_sum = predicted_sum = true_sum = 0

    # Process each instance
    for true, predicted in zip(true_labels, predicted_labels):
        # Extend the sets with ancestors
        extended_true = true.union(
            *[ancestors[label] for label in true if label in ancestors]
        )
        extended_predicted = predicted.union(
            *[ancestors[label] for label in predicted if label in ancestors]
        )

        # Update counters
        true_positive_sum += len(extended_true.intersection(extended_predicted))
        predicted_sum += len(extended_predicted)
        true_sum += len(extended_true)

    # Calculate hierarchical precision and recall
    hP = true_positive_sum / predicted_sum if predicted_sum else 0
    hR = true_positive_sum / true_sum if true_sum else 0

    # Calculate hierarchical F-measure
    hF = ((beta**2 + 1) * hP * hR) / (beta**2 * hP + hR) if (hP + hR) else 0

    return hP, hR, hF


# Example usage:
true_labels = [{"G"}]  # The true class for the instance
predicted_labels = [{"F"}]  # The predicted class for the instance
ancestors = {  # The ancestors for each class, excluding the root
    "G": {"B", "C", "E"},
    "F": {"C"},
}

# Calculate hierarchical measures
hP, hR, hF = hierarchical_precision_recall_fmeasure(
    true_labels, predicted_labels, ancestors
)
print(f"Hierarchical Precision (hP): {hP}")
print(f"Hierarchical Recall (hR): {hR}")
print(f"Hierarchical F-measure (hF): {hF}")