Skip to content
BENIBIX Learn

Decision Threshold

AI📏 Evaluation & BenchmarksClassification Metrics

Also known as Classification Threshold, Probability Threshold

The cut-off that turns a model's score into yes or no. Moving it trades misses for false alarms.

Think of it like

It is like the height line at a fair ride. If you are above the line, you can go on.

Example

Lucy's fraud model blocks a payment when its score is above 0.5. Too many good payments get blocked, so she raises the line to 0.8.

Why it matters

The same model can be careful or bold just by moving one number. Precision and recall both change with it.

Definition

A decision threshold is the cut-off that turns a model's score into a yes or a no. Many models give a number from 0 to 1, such as how likely an email is spam. If the number is above the line, the answer is yes. The usual line is 0.5, but you can move it. A higher line means fewer false alarms but more misses. A lower line catches more, but it cries wolf more often.
Move the line up and fewer emails are called spam. Fewer false alarms, but some spam slips through.

Common mistake

Leaving the line at 0.5 without asking why. The right line depends on which mistake costs more.

Key takeaways

  1. 1
    Score to answer. A number becomes a yes or a no.
  2. 2
    Move the line. Higher means fewer false alarms.
  3. 3
    Cost decides. Pick it by which mistake hurts more.