A Machine learning approach to fraudulent payment detection of the payment aggregator business model in Sri Lanka

Abstract

When a payment aggregator is accepting payments on behalf of the merchants as a financial technology service provider, it is possible for a portion of those payments to be fraudulent payments. In order to detect fraudulent payments and avoid related losses, it is important to use an algorithm-based model which adapts to the changing circumstances instead of having fixed rules. This research uses machine learning concepts in order to detect a given card-not-present online transaction being a suspicious-for-fraud payment in the context of a Central Bank approved payment aggregator business model in Sri Lanka. Further to that, this research also investigates the conditions which are highly influential in deciding whether a given payment is suspicious for fraud under the payment aggregator model. Under machine learning, classification approach is used, as the dataset used is categorized as fraud and not fraud. The attributes related to the payment data are different given the context and feature engineering was required to obtain a meaningful outcome. It was discovered that the conditions which influence a payment to become suspicious for fraud are the name registered with the payment method being different to the name of the actual payee who made that particular transaction and the originating country of the transaction being different to the name of the country entered by the payee who made that particular transaction. Fourteen different supervised learning algorithms were tested on a payment dataset and were evaluated based on the accuracy of the predicted class label. As part of the outcome of this study, decision tree algorithm was identified as the most effective algorithm with the highest prediction accuracy and a model was built and saved using PyCaret for future use.