Anomaly Detection in Enterprise Payment Systems: An Ensemble Machine Learning Approach

With the exponential growth of digital transactions, ensuring the integrity and authenticity of payment systems has become imperative. This paper investigates the effectiveness of machine-learning techniques in detecting anomalous patterns in large-scale payment datasets. Ensemble methods are widely used in the field of anomaly detection in enterprise systems to improve the accuracy and robustness of these systems. Anomaly detection aims to detect abnormal patterns that deviate from the rest of the data and are referred to as anomalies or outliers. With millions of services or sub-systems to monitor such as e-commerce platforms and governmental portals, our study focuses on using forecasting methods to develop a model that can be used in these enterprise systems to avoid huge financial impacts, bad reputation, and customer dissatisfaction. Our methodology combines multiple time series methods such as Seasonal Autoregressive integrated moving average (SARIMAX) and Facebook-Prophet and SVM to create a more robust and accurate ensemble model for anomaly detection. Anomaly detection can help highlight where exactly an incident is occurring. This proactive detection greatly improves the root cause analysis of the problem and has a positive impact on business continuity. The three different types of anomalies can occur in the datasets of pointers, conditional, and collective or accumulative anomalies. The main approaches to solve anomaly detection problems are either rule-based or machine learning approaches. In this paper, we focus on the machine learning approach as it is more reliable and effective as it complements the rule-based human capabilities with the machine learning and artificial intelligence capabilities. Three widely used forecast models SARIMAX, Facebook-Prophet and SVM are compared and analysed for the payment transactions. For the evaluated performance SVM model is best performed with R squared accuracy values of 80.7%. Overall, the results demonstrated that the SVM method can provide better performance than SARIMAX and Prophet methods for payment transactions data.