Balancing Durability, Performance, and Interpretability in Unbalanced Data as Fraud Detection

Abstract

The problem of fraud detection is one of the most discussed topics in the field of machine learning. This study addresses four key areas essential for a fraud detection platform: prediction accuracy in imbalanced datasets, interpretability of predictions, deployment and sustainability of the platform, monetary costs associated with model errors. To tackle these issues, we first conducted extensive research in the field, then proposed and evaluated our solutions. We introduce methods such as using a WCGAN (Wasserstein Conditional Generative Adversarial Network) for sampling or cost-sensitive learning with new models like Light Gradient Boosting, employing interpretable models like Explainable Boosting, deploying and automating training processes with Kubernetes and Kubeflow, and utilizing approaches like thresholding or tuning metrics that account for monetary costs. Each of these solutions shows promising results and improves upon existing research in the field.