• From one-class to two-class classification by incorporating expert knowledge

      Oosterlinck, Dieter; Benoît, Dries; Baecke, Philippe (2018)
      In certain business cases the aim is to identify observations that deviate from an identified normal behaviour. It is often the case that only instances of the normal class are known, whereas so called novelties are undiscovered. Novelty detection or anomaly detection approaches usually apply methods from the field of outlier detection. However, anomalies are not always outliers and outliers are not always anomalies. The standard one-class classification approaches therefore underperform in many real business cases. Drawing upon literature about incorporating expert knowledge,we come up with a new method that significantly improves the predictive performance of a one-class model. Combining the available data and expert knowledge about potential anomalies enables us to create synthetic novelties. The latter are incorporated into a standard two-class predictive model. Based on a telco dataset, we prove that our synthetic two-class model clearly outperforms a standard one-class model on the synthetic dataset. In a next step the model was applied to real data. Top identified novelties were manually checked by experts. The results indicate that incorporating expert knowledge to transform a one-class problem into a two-class problem is a valuable method.