core Estimated learning time: 4 h

5.16 Anomaly detection

You can flag rare events without a labelled dataset.

Before:03. Data Handling & AnalysisUnlocks:06. Deep Learning10. Production & MLOps13. Capstones, Portfolio & Interviews

Anomaly detection flags rare events without labelled examples — statistical thresholds, Isolation Forest, autoencoder reconstruction error — and it powers fraud screens, fault detection and monitoring. It sits here because it composes the unsupervised ideas into a deployable job. The hard part is evaluation: with almost no true positives, accuracy is meaningless, precision is fragile, and knowing which numbers survive rarity is half the skill.

Work through these

  • Statistical thresholds and robust z-scores

    Simple thresholds on how far a value sits from typical, using measures that are not themselves distorted by the outliers. It is the baseline every more elaborate method must beat.

  • Isolation Forest and One-Class SVM

    Two methods designed for the unlabelled case, one isolating unusual points by how quickly they can be separated, the other learning a boundary around normal data. Both are standard first choices.

  • Autoencoder reconstruction error

    Training a network to reconstruct normal data and flagging what it reconstructs badly. It suits high-dimensional data where simpler distances stop working.

  • Evaluating detectors when positives are rare

    With very few positives, ordinary metrics become meaningless and evaluation needs care about what a false alarm actually costs. This is where most anomaly projects go wrong.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.