Deep fraud. A fraud intention recognition framework in public transport context using a deep-learning approach

In this paper, we present a framework for fraud intention recognition of public transport bus operators based on a deep learning approach using a stack of denoising and sparse autoencoders. Bus operator's fraud is a common problem in passenger transport organizations with cash payment method and tickets as proof of passenger on board. Fraud detection could be considered as an anomaly detection problem, and several techniques in this area have been applied to it. Our approach is an architecture based on a sort of Stacked Autoencoders with a Softmax Classification Layer and a normalization pre-phase using word2vec and a Denoising Autoencoder. This framework permits to recognize the fraud intention in the operators delivers account time, based on date, route, count of passengers on board, among other features. In this framework, the fraud intention is modeled as a binary classification problem, there is fraud intention, or there isn't. We compare our approach with another nondeep state of the art classification approaches like Logistic Regression, SVM, Decision Trees, KNN and Adaboost ensemble method, obtaining a superior performance on intention recognition, with an accuracy of 87.3 percent and an F1 score of 0.752.