Leveraging Big Data and AI/ML for Fraud Detection in Retail Transactions

Abstract

Fraud is an intentional misrepresentation or deception that causes another person to act to his or her detriment. Fraud detection is a critical problem faced by many organizations, particularly in industries such as e-commerce, telecommunications, banking, and insurance. New developments in smart devices, new retail transactions, and the massive growth of these devices have introduced many new opportunities for fraudsters as well as challenges for organizations trying to combat fraud. Retail transactions containing personally identifiable information are now more prone to privacy concerns, data breaches, and fraud than ever before. Legacy fraud detection techniques of manually crafting rules and employing rules without real scientific analysis are ineffective against new retail transactions. Big data technologies provide industry-leading speed, performance, and scale insight. AI/ML models are showing superior performance for fraud detection tasks due to new developments in computer science and the availability of data streams. Retail transactions in the form of large-scale datasets generated in real-time can be tackled using big data technologies and in-database AI/ML. A novel and effective AI/ML-based fraud detection approach is proposed that employs big data and AI/ML concepts to safeguard organizations from fraudsters. Leveraging the distributed in-database capabilities of big data technology, fault-tolerant massive datasets can be processed in parallel and analyzed in seconds using AI/ML algorithms. The ingestion of massive datasets within the architecture leads to AI/ML models being employed in the big data technology environment, significantly enhancing speed, performance, and reliability. The approach consists of four models: 1. Data Sources - a detailed overview of data sources is provided, including simulated datasets with retail transactions. 2. Data Processing and AI/ML Architecture - a description of preprocessing tables and features for deployment in big data technologies and design of AI/ML architecture for modeling and deployment, where AI/ML models are embedded in SQL, fed by processed tables in data lakes, and scoring tables are created with predictions, is provided. 3. Results - the performance of the approach with four AI/ML models employed on simulated datasets is evaluated. 4. Conclusions and Future Work - a summary of the approach is provided together with the performance of the designed models, and several future work ideas are introduced. The approach is novel in its design and implementation using new technologies and datasets while focusing on retail transactions, and it is valid since robust performance is shown with diverse AI/ML models.