Active weakly-supervised anomaly detection for relational data
Anomaly detection is ubiquitous in today’s society: banks and insurance companies detect fraud by finding anomalies in financial transactions; machines at intensive care continuously monitor patients to notify staff whenever an anomaly occurs; and so on. Generally, an anomaly is a deviation from normal behaviour. Developing traditional anomaly detectors without machine learning is costly and time-consuming. It is very hard to define precisely what constitutes an anomaly. This might be feasible for small systems and simple anomalies but is infeasible for large systems or complex anomalies.
There are three problems with current machine learning anomaly detection systems. (1) They make strong assumptions on what constitutes an anomaly (e.g. infrequent behaviour is anomalous). In real world scenarios these assumptions are often violated. (2) Current systems are limited in the kind of information they can use to learn an anomaly detector. (3) Often the different observations are assumed independent, if this is not the case the detector might not detect certain anomalies.The goal of this project is to design a novel anomaly detection algorithm that solves the three problems listed above. Our approach will be weakly supervised (i.e. allows to integrate more general forms of information), active (i.e. makes efficient usage of a domain expert’s knowledge) and employs relational data to capture the dependencies between observations.