Anomaly detection
Methods that flag unusual observations relative to a defined reference or a learned pattern.
When is an observation unusual?
Anomaly detection flags observations that depart unusually from a defined reference or a learned pattern. What counts as normal depends on the selected data, features and method. Some approaches look for outliers in existing data; others assess new observations against a data basis considered normal. A flag provides a reason to investigate. It does not explain the cause of the deviation or, by itself, establish error or fraud.
In a fictional service business, a system is to flag an unusual rise in recorded requests. The team plans cases involving actual additional customer requests and old cases imported together in a batch. Usual differences between weekdays are also to be considered. A responsible person checks whether more work has arisen or recording has changed the count. These situations require different responses, such as capacity planning or correcting the analysis. False warnings and missed unusual cases are recorded during the trial; a high count alone does not determine any of those responses.
Sources
Related terms
- Machine learningModels learn patterns from data and apply them to new inputs.
- AI inferenceApplying an already trained model to inputs to obtain a result.
- LLM evaluationLLM evaluation tests model outputs against defined tasks and criteria. One good example does not establish overall quality.
- Forecasting modelA model that uses available information to estimate a future quantity for a defined period.
- Unsupervised learningA learning method that looks for patterns in data without predefined target answers for the particular training task.
- Model cardA model card documents an AI model’s purpose, foundations, evaluation conditions and known limitations for a particular version.
Articles on this topic
AI & Automation with ASES
AI assistants and automated workflows