liveBook · Manning

Precision

As explained in Machine Learning System Design

The proportion of predicted positive cases that are actually positive.

A fraud model flags transactions for review. Precision asks: of the transactions flagged as fraud, how many really were fraud? It is true positives divided by all predicted positives: TP / (TP + FP). section 5.2

Same recall, different alert quality

Suppose 100 transactions are fraudulent and two models each catch 80. Model A also flags 20 legitimate transactions, giving 80% precision. Model B flags 920 legitimate transactions, giving 8% precision. Both have 80% recall, but their alert queues differ sharply. section 5.2

Precision also changes when the mix of fraud and legitimate transactions changes, even if a model’s detection and false-positive rates stay fixed. Compare it with recall and specificity when fraud prevalence shifts; no single metric is ideal for every goal. section 5.2

How Machine Learning System Design covers it

How deep it goes
Defines the ratio, works through false-positive comparisons, and cautions that changing class balance affects it.
Examples
Compares fraud models with equal recall but 20 versus 920 false positives.
What it assumes
Assumes binary fraud labels and uses true-positive and false-positive counts to calculate the metric.
What you can do after
Evaluate/understand it by checking whether changing fraud prevalence or false positives explain a reported precision score.
How it connects
Relates precision to recall, specificity, false-positive rate, and class imbalance.
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 20 books explain precision →

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.