In predictive analytics, the aim is to build an analytical model predicting a target measure of interest (Baesens 2014; Duda et al. 2001; Flach 2012; Han and Kamber 2001; Hastie et al. 2001; Tan et al. 2006). The target is then typically used to steer the learning process during an optimization procedure. Two types of predictive analytics can be distinguished depending on the measurement level of the target: regression and classification. In regression, the target variable is continuous and varies along a predefined interval. This interval can be limited (e.g., between 0 and 1) or unlimited (e.g., between 0 and infinity). A typical example in a fraud detection setting is predicting the amount of fraud. In classification, the target is categorical which means that it can only take on a limited set of predefined values. In binary classification, only two classes are considered (e.g., fraud versus no-fraud) whereas in multiclass classification, the target can belong to more than two classes (e.g., severe fraud, medium fraud, no fraud).
In fraud detection, both classification and regression models can be used simultaneously. Consider, for example, an insurance fraud setting. The expected loss due to fraud can be calculated as follows
where PF represents the probability of fraud and LGF the loss given fraud. The latter ...