Fraud detection by analyzing human behavior apply machine learning techniques.

Abstract

Currently, we are witnessing a sudden increase in fraud cases around the world, which makes it necessary to adopt proactive strategies to detect indications or suspicions before the crime materializes. Fortunately, the continuous development of computer technologies provides favorable opportunities to combat and mitigate this problem. Text mining and topic modeling are practical tools that, together with theories focused on analyzing this phenomenon, can help identify fraud-related issues and discover fraudulent activities to make accurate decisions. Since tangible information on fraud is limited, several synthetic datasets with phrases based on the fraud triangle theory were generated, which was used to develop the proposed model. To recognize patterns in documents, unsupervised topic modeling techniques were applied, followed by a semi-supervised approach that outperformed unsupervised models and provided superior interpretability; This allowed for establishing a relationship between the resulting topics and the vertices of the fraud triangle and obtaining high probabilities that a document belongs to a specific topic. These probabilities were used to train classification algorithms and predict behavior suspected of fraud, with promising results. Evaluating the model’s applicability to all the generated datasets allowed us to determine that the model was generalizable and, therefore, useful for fraud detection.