Anomaly Detection is looking for features which are ‘odd’ or ‘out of the norm’ and which might indicate e.g. fraudulent behavior or alert to something which might go wrong in a mechanical system even though there is no specific trigger threshold for it.
Anomaly Detection can be ‘supervised’ or ‘unsupervised’ (or indeed semi-supervised). Arguably, unsupervised is more powerful it is suggesting signals which look odd without necessarily having seen them before and having a reference ‘decision marker’ for those features. It can be supervised (or trained) afterwards to avoid ‘information overload’ i.e. too many alarms and alerts. Potentially it can do this by being combined with other techniques like clustering to ‘mute’ classes of anomalies or features within the data which should not trigger an alarm. Use cases might be looking for fraud and error, or early indications of issues occurring in a live system or process e.g. manufacturing or operating machinery. There are many algorithms for this category, and many proprietary algorithms and suites e.g. coming out of the finance, fraud and IT security industries.