Critical Analysis of Machine Learning Based Approaches for Fraud Detection in Financial Transactions

Abstract

Fraud has become a trillion-dollar industry today. Some finance companies have separate domain expert teams and data scientists who are working on identifying fraudulent activities. Data Scientists often use complex statistical models to identify frauds. However, there are many disadvantages to this approach. Fraud detection is not real-time and therefore, in many cases fraudulent activities are identified only after the actual fraud has happened. These methodologies are prone to human errors. In addition, it requires expensive, highly skilled domain expert teams and data scientists. Nevertheless, the accuracy of manual fraud detection methodologies are low and due to that, it is very difficult to handle large volumes of data. More often, it requires time-consuming investigations into the other transactions related to the fraudulent activity in order to identify fraudulent activity patterns. Finance companies are not getting adequate return of interest (ROI) despite the resources and money spent on these traditional methodologies. Most of the traditional fraud detection methodologies focused on discrete data points. (User accounts, IP addresses devices, etc...) However, these methodologies are no longer sufficient for today's needs. As fraudsters and hackers are using more advanced and cutting edge techniques to mask their fraudulent activities even from the sharpest eyes. These methodologies can only detect known types of attacks, therefore an analytical approach is required to address these drawbacks of the traditional methodologies.