Aller au contenu
Atlas de l'IA

Détection d’anomalies

Repérer les quelques cas étranges parmi une masse de données normales

Données et documentsIntermédiaire #40
entréeTableauTexte

Le texte intégral est présenté en anglais ; le titre et le résumé sont localisés.

CE QUE DÉSIGNE CETTE CAPACITÉ

Takes data records or time series and outputs how far each deviates from the norm, or a flag. It usually works with few or no labelled anomalies, modelling what normal looks like and calling deviations anomalous. Unlike text classification the anomaly classes are not fixed, and in training you often do not know what an anomaly looks like at all.

Comment c'est fait

Unsupervised routes model normal data: autoencoders flag large reconstruction errors, isolation forests treat easily isolated points as anomalous, and statistical methods fit a distribution and score low-probability points. Time series often use a predictive framing, taking the residual between actual and forecast values as the score. With a few labels, semi-supervised or contrastive learning takes over, and cost-sensitive thresholds balance misses against false alarms.

Produits représentatifs

3

Organisations concernées

Usages typiques

  • Early warning for equipment and production-line faults
  • Spotting oddity in transactions and money laundering
  • Intrusion and abnormal-traffic monitoring
  • Quality sampling and data-entry error catching

Comment on l'évalue

AUROC
Threshold-free ranking quality, usable on heavily imbalanced data
PR-AUC
More informative than AUROC when anomalies are extremely rare
Detection delay
Time from onset to alert, key in real-time settings

Limites et points difficiles

  • With so few anomalies a small threshold change spikes false alarms, and operators quickly go numb to them
  • Concept drift marks normal behaviour as anomalous, and seasonality or business change demands continual recalibration
  • In high dimensions with correlated features distance-based measures break down, diluting the gaps between normal points

Concepts sous-jacents