An Experimental Analysis of Fraud Detection Methods in Enterprise Telecommunication Data using Unsupervised Outlier Ensembles

This work uses outlier ensembles to detect fraudulent calls in telephone communication logs made on the network of POST Luxembourg. Outlier detection on high-dimensional data is challenging and developing an approach which is robust enough is of paramount importance to automatically identify unexpected events. For use in real-world business applications it is important to obtain a robust detection method, i.e. a method that can perform well on different types of data, to ensure that the method will not impact that business in unexpected ways. Many factors affect the robustness of an outlier detection approach and this experimental analysis exposes these factors in the context of outlier ensembles using feature bagging. Real-world problems demand knowledge about possible candidate approaches that address the problem, and decide for the best performing method using a train-test split of labeled data. In the unsupervised setup the knowledge about performance is missing during the learning phase thus is difficult to decide during that phase. Hence, in this setup it is important to know about how the performance is affected before the learning phase. Hence, this analysis demonstrates that despite the collective power of outlier ensembles they are still affected by i) data normalization schemes, ii) combination functions iii) outlier detection algorithms.