False Positive

A model can be “wrong” in two very different ways: it can miss something that’s truly there, or it can raise an alarm when nothing is there. A false positive is the second kind of mistake—one that feels like a false alarm.

What it means (in classification)

In a binary classifier (yes/no, positive/negative), a false positive happens when the model predicts the positive class, but the true label is actually negative. In a confusion matrix, it’s the count in the “predicted positive, actually negative” cell. For example, in spam detection, a legitimate email marked as spam is a false positive; in fraud detection, a normal transaction flagged as fraud is a false positive.

Why it matters in practice

False positives have real costs, and those costs depend on the application:

• Medical screening: a healthy patient flagged as sick can trigger anxiety and unnecessary follow-up tests.
• Credit scoring: a reliable borrower incorrectly rejected is lost business and potential fairness concerns.
• Security/fraud: too many false alarms overwhelm investigators and slow response to real threats.

Because of this, teams track metrics tied directly to false positives, like precision (how many predicted positives are truly positive) and the false positive rate (FPR = FP / (FP + TN)).

How models trade false positives vs. false negatives

Many classifiers output a probability score, then apply a decision threshold (e.g., predict “positive” if probability ≥ 0.5). Lowering the threshold usually increases true positives but also increases false positives. Tools like the ROC curve and precision-recall curve visualize this trade-off; in scikit-learn you’ll see functions like roc_curve and precision_recall_curve used for exactly this diagnostic work.