Detect unusual observations, choose thresholds, and assess the resulting review workload. Separate novelty detection from outlier detection and evaluate false p
Last reviewed: 2026-10-03
Anomaly detection identifies observations that differ from a reference pattern. Outlier detection allows unusual observations in the training set; novelty detection typically learns a reference from data treated as normal and scores new observations. These assumptions affect which methods and evaluation procedures are appropriate.
For fraud or security triage, an anomaly is a reason to investigate, not a finding of wrongdoing. Tune thresholds against a review budget and inspect false positives. Changes in normal behavior, missing data, and data-collection errors can all produce unusual scores.
No. It indicates departure from the model's reference pattern. Use corroborating evidence and an appropriate review process before acting on an individual record.