Development of users' call profiles using unsupervised random forest

The aim of this paper is to detect fraud in telecommunications data which consists of millions of call records generated each day. The fraud detection is implemented via the construction of user call profiles using the calls detail records (CDR) data. This paper attempts to investigate the reliability of the unsupervised Random Forest method in building the profiles using its variable importance measure. Four different simulation scenarios, using different number of variable selection in each node of the tree, are performed.