Which metric evaluates classifier performance across varying thresholds?

Which metric evaluates classifier performance across varying thresholds?

• A. Recall
• B. ROC - AUC (Receiver Operating Characteristic - Area Under Curve)
• C. Specificity
• D. Precision

Answer: (B)

Evaluating how a classifier performs across different thresholds focuses on how well it separates positives from negatives no matter where you draw the cutoff. The ROC curve does exactly this by plotting, for many thresholds, the true positive rate against the false positive rate. It shows the tradeoff between catching positives and misclassifying negatives as positives as you vary the threshold. The Area Under the Curve then compresses that whole curve into a single number, giving an overall measure of the model’s ability to discriminate across all thresholds. This matters beyond individual cutoff choices because it captures ranking capability: a model with a higher AUC tends to assign higher scores to positives than to negatives across the board. In contrast, recall, specificity, and precision are all tied to a specific threshold and don’t summarize performance across the full range of thresholds. Recall is the proportion of positives found at one cutoff, specificity is the proportion of negatives correctly identified at that cutoff, and precision depends on the chosen threshold and the prevalence of the positive class. So the metric that evaluates classifier performance across varying thresholds is the ROC curve with its AUC. It provides an aggregate sense of discriminative ability across all possible thresholds.

Evaluating how a classifier performs across different thresholds focuses on how well it separates positives from negatives no matter where you draw the cutoff. The ROC curve does exactly this by plotting, for many thresholds, the true positive rate against the false positive rate. It shows the tradeoff between catching positives and misclassifying negatives as positives as you vary the threshold. The Area Under the Curve then compresses that whole curve into a single number, giving an overall measure of the model’s ability to discriminate across all thresholds.

This matters beyond individual cutoff choices because it captures ranking capability: a model with a higher AUC tends to assign higher scores to positives than to negatives across the board. In contrast, recall, specificity, and precision are all tied to a specific threshold and don’t summarize performance across the full range of thresholds. Recall is the proportion of positives found at one cutoff, specificity is the proportion of negatives correctly identified at that cutoff, and precision depends on the chosen threshold and the prevalence of the positive class.

So the metric that evaluates classifier performance across varying thresholds is the ROC curve with its AUC. It provides an aggregate sense of discriminative ability across all possible thresholds.