Neural Network Lexicon

AUC

Short Definition

AUC measures how well a model separates positive and negative classes.

Definition

Area Under the Curve (AUC) refers to the area under the ROC curve. It represents the probability that a randomly chosen positive example is ranked higher than a randomly chosen negative example by the model.

AUC summarizes the ROC curve into a single scalar value between 0 and 1.

Why It Matters

AUC provides a threshold-independent measure of a model’s ranking ability. It is particularly useful when comparing models or when class distributions vary between datasets.

An AUC of:

  • 0.5 indicates random guessing
  • 1.0 indicates perfect separation

How It Works (Conceptually)

  • The ROC curve is constructed across thresholds
  • The area under that curve is computed
  • Larger area indicates better separability

AUC evaluates ranking quality, not prediction calibration.

Interpretation

AUC ≈ 0.5 → no discriminative power
AUC > 0.7 → useful discrimination
AUC > 0.9 → very strong discrimination

Minimal Python Example

Python
auc = compute_auc(y_true, y_scores)

Common Pitfalls

  • Using AUC as the only evaluation metric
  • Assuming high AUC implies high precision or recall
  • Ignoring performance at the actual operating threshold
  • Comparing AUC across unrelated tasks

Related Concepts

  • ROC Curve
  • Precision–Recall Curve
  • Evaluation Metrics
  • Model Confidence
  • Calibration