Short Definition

Decision thresholding controls how prediction scores are converted into class labels.

Definition

Decision thresholding is the process of selecting a cutoff value that determines when a model’s predicted score or probability is assigned to a particular class. In binary classification, predictions above the threshold are labeled positive, and those below are labeled negative.

The default threshold is often 0.5, but this is rarely optimal.

Why It Matters

Changing the decision threshold alters the balance between false positives and false negatives. The optimal threshold depends on the application, error costs, and class distribution.

Thresholding allows models to be tuned for specific operational requirements.

How It Works (Conceptually)

• The model outputs a score or probability
• A threshold is applied to that score
• Predictions are converted into labels
• Metrics change as the threshold changes

Thresholding defines the operating point of a model.

Minimal Python Example

prediction = score >= threshold

Common Pitfalls

• Using default thresholds blindly
• Optimizing thresholds on test data
• Ignoring calibration
• Evaluating metrics without specifying the threshold

Related Concepts

• Precision
• Recall
• ROC Curve
• Model Confidence
• Calibration