From one-class to two-class classification by incorporating expert knowledge: Novelty detection in human behaviour

One-class classification is the standard procedure for novelty detection. Novelty detection aims to identify observations that deviate from a determined normal behaviour. Only instances of one class are known, whereas so called novelties are unlabelled. Traditional novelty detection applies methods from the field of outlier detection. These standard one-class classification approaches have limited performance in many real business cases. The traditional techniques are mainly developed for industrial problems such as machine condition monitoring. When applying these to human behaviour, the performance drops significantly. This paper proposes a method that improves existing approaches by creating semi-synthetic novelties in order to have labelled data for the two classes. Expert knowledge is incorporated in the initial phase of this data generation process. The method was deployed on a real-life test case where the goal was to detect fraudulent subscriptions to a telecom family plan. This research demonstrates that the two-class expert model outperforms a one-class model on the semi-synthetic dataset. In a next step the model was validated on a real dataset. A fraud detection team of the company manually checked the top predicted novelties. The results show that incorporating expert knowledge to transform a one-class problem into a two-class problem is a valuable method.