Anomalies, or outliers, occur in a wide range of applications, and their detection may be significant, for example, for diagnosing diseases, tracking criminal actions, controlling production pipelines, or uncovering operator network changes. In general, the anomalies are contained in unlabelled data, and various unsupervised learning algorithms have been developed for uncovering deviant observations that do not conform to the expectations. However, due to the lack of ground-truth labels, it remains challenging to evaluate how well the anomalies are captured by the model. This presentation discusses the challenges in evaluating whether the machine learning model finds the most significant anomalies focusing on a case study on operator network data.
- It is important to characterise a normal observation, to be able to decide what is not normal
- It is generally beneficial to examine the data using various anomaly detection approaches
- Simulation studies can assist in comparing the capabilities of the different methods