Metrics • Intermediate

True Positive Rate vs False Positive Rate

A side-by-side comparison of True Positive Rate and False Positive Rate. Understand how detection of actual positives differs from mistaken positive predictions among actual negatives.

Quick Verdict: Use True Positive Rate to measure how many actual positives are found; use False Positive Rate to measure how many actual negatives are incorrectly flagged.

At a Glance

True Positive Rate

True Positive Rate describes proportion of actual positive cases that a model correctly identifies as positive.

Key Characteristics

• Proportion of actual positive cases correctly identified
• Also known as recall or sensitivity
• Measures detection effectiveness for the target class

Watch Out For

• High TPR may come with more false positives
• Should be interpreted with class balance and threshold choices

Context: Most relevant when missing positive cases is costly or safety-critical.

False Positive Rate

False Positive Rate describes proportion of actual negative examples that a model incorrectly predicts as positive.

Key Characteristics

• Proportion of actual negative examples incorrectly predicted as positive
• Common in binary classification evaluation
• Forms the x-axis of the ROC curve

Watch Out For

• Low FPR does not guarantee high detection of positives
• Can appear small in highly imbalanced datasets while still creating operational burden

Context: Most relevant when false alarms, unnecessary interventions, or wrongly flagged users carry cost or harm.

Key Differences

Aspect	True Positive Rate	False Positive Rate
What it measures	TPR measures the share of actual positives that the classifier correctly identifies.	FPR measures the share of actual negatives that the classifier incorrectly labels as positive.
Best use case	Best when the core question is whether the system catches the target condition.	Best when the core question is how often the system raises false alarms.
Failure mode	A low TPR means the model misses too many actual positive cases.	A high FPR means the model incorrectly flags too many negative cases.
Threshold sensitivity	TPR often increases when thresholds are lowered to catch more positives.	FPR often increases when thresholds are lowered because more negatives are also flagged.
Common mistake	Reporting high TPR without showing the false-positive trade-off.	Reporting low FPR without showing whether the model still detects enough positives.

Caesar AI Note

In practice, TPR and FPR should usually be reviewed together. A model can look strong on one metric while creating unacceptable risk through the other.

Notes

Common Mistakes

Calling TPR accuracy.

Ignoring threshold settings when comparing models.

Using TPR without reviewing false positives.

Using FPR without considering missed positive cases.

When to Use Each

true-positive-rate

Use True Positive Rate when the evaluation question is how effectively the model detects actual positive cases. It is especially important for screening, safety, fraud, health, or risk tasks where missed positives create harm.

false-positive-rate

Use False Positive Rate when the evaluation question is how often the model wrongly flags negative cases. It is important where false alarms create cost, discrimination, unnecessary review, or user harm.

Compliance Note

Metric selection affects AI validation reports, risk acceptance, and claims made to users or regulators. NIST AI RMF and EU AI Act evidence should show threshold choices, error trade-offs, and impact analysis rather than isolated metrics.

FAQ

Is true positive rate the same as recall?+

Yes. True positive rate is also known as recall or sensitivity and measures the share of actual positives correctly identified.

Why is false positive rate used in ROC curves?+

The false positive rate forms the x-axis of the ROC curve, while the true positive rate forms the y-axis. Together they show the detection and false-alarm trade-off across thresholds.

Which metric matters more?+

That depends on the harm model. If missed positives are worse, TPR may dominate; if false alarms are worse, FPR may dominate, but both should be documented.