Precision

What it is

When a model classifies items, such as flagging toxic comments or identifying objects in an image, it outputs a probability score. You typically apply a classification threshold to turn those scores into binary yes or no decisions. Precision measures how many of your yes decisions were accurate compared to the total number of yes decisions made. If you label 100 emails as spam and 90 are truly spam, your precision is 0.90.

Why it matters

You care about precision when the cost of a false positive is high. For example, if your system automatically bans users for policy violations, low precision means you are banning innocent users, which leads to support tickets and user churn. If you ignore precision, you might prioritize a high volume of catches while accidentally damaging the user experience with incorrect actions.

In practice

You usually control precision by adjusting the decision threshold in your inference logic. A higher threshold makes the model more conservative, which typically increases precision at the expense of missing more valid cases. You monitor this in production by comparing model predictions against human-verified logs or a gold standard test set.

The tradeoff

Precision and recall represent a zero-sum trade-off: tightening criteria to improve precision almost always causes your system to miss more legitimate targets, thereby lowering recall.

Where it appears

Research summaries that use Precision, each linked to its source paper.

• Improving Knowledge Graph Question Answering Accuracy
  Constrained Entity Selection under Partial Knowledge for LLM-Based Knowledge Graph QA

  The paper introduces a constraint-based verification method that filters LLM responses against knowledge graph logic to significantly improve answer precision without sacrificing recall.

• Automated Security Annotation for Business Processes
  A Hybrid LLM-Based Framework for Automated Security Annotation Generation in Business Process Models

  The paper introduces a hybrid LLM-based framework that automates the generation of SecBPMN2 security annotations from natural-language specifications to improve process model accuracy.

• Optimized Compression for Image Models
  HAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression

  The paper introduces HAMP-LIC, a Hessian-aware mixed-precision quantization method that shrinks learned image compression models while preserving image quality and eliminating cross-platform decoding mismatches.