liveBook · Manning

False-positive rate

As explained in Machine Learning System Design

The proportion of actual negative cases that a model incorrectly predicts as positive.

In fraud detection, a false positive is a genuine case flagged as fraud. False-positive rate (FPR) is the share of actual negative cases flagged positive: FP / (FP + TN). A small rate can still create many incorrect alerts when the volume is large. section 5.2

Compare the alerts, not just the rates

Two fraud models achieve the same recall, but model A has an FPR of 0.0002 and model B an FPR of 0.0092—46 times higher. Of model B’s 1,000 fraud predictions, 920 are wrong. Its precision of 0.08 makes the problem easier to see. section 5.2

Choose an acceptable FPR for the expected volume, then consider how much fraud the model catches at that limit. In a high-volume spam and fraud system, a specificity target of 0.999999 set the false-positive limit; recall was maximized within it. Precision can change when class balance shifts even if FPR and recall stay unchanged. section 5.2

How Machine Learning System Design covers it

How deep it goes
Defines the calculation and works through model comparisons, without deriving a statistical decision rule.
Examples
Compares two fraud models and describes a high-volume spam and fraud system with a specificity target.
What you can do after
Evaluate/understand model tradeoffs by setting a tolerable false-positive rate and checking recall and precision.
How it connects
Relates false-positive rate to specificity, recall, precision, and class balance.
Cover of Machine Learning System Design
liveBookmanning.com
Read Machine Learning System Design on liveBook Get the book at manning.com

Also explained in

Compare how 9 books explain false-positive rate →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machine Learning System Design. The section links point to the parts of the book it draws on.