ROC AUC measures a classifier's ability to separate classes across every possible threshold by summarizing its receiver operating characteristic curve into a single number.
ROC AUC is the area under the receiver operating characteristic curve. The ROC curve plots true positive rate against false positive rate as the classification threshold moves from always positive to always negative. An AUC of 1.0 is perfect separation; 0.5 is random guessing.
That single number summarizes threshold-independent performance. Accuracy and F1 score depend on one chosen threshold; AUC evaluates the model across all possible trade-offs. It is especially helpful when the costs of false positives and false negatives shift. The curve shows whether the model can achieve high recall without sacrificing too much precision.
Think of it like this. Think of a medical test where you can adjust how strict the diagnosis is. The ROC curve shows every possible strictness level, and AUC summarizes whether the test is good at separating sick from healthy patients overall.
The model outputs a score for every example. Sorting examples by score and sweeping a threshold from high to low traces the ROC curve. At each threshold, the true positive rate and false positive rate are recorded. The area under that curve is computed using the trapezoid rule, producing a value between 0 and 1.
"High AUC means the model is ready for production." AUC ignores calibration and class imbalance. "AUC is the only metric I need." It hides the shape of the curve, which matters for cost-sensitive decisions. "AUC below 0.5 means the model is inverted." It can mean that, or it can mean the score ordering is wrong for the chosen positive class.
Threshold-independent and interpretable, but sensitive to class imbalance and can be misleading when the positive class is rare. Best used alongside PR AUC and business-cost analysis for imbalanced problems.