Utilization of encoding, early stopping, hyper parameter tuning, and machine learning models for bank fraud detection

Abstract

An effective fraud detection system must protect millions of clients for a secure banking system, which can be achieved using machine learning and AI. In this article, authors have applied four supervised machine learning models: k-nearest neighbors (KNN), random forest (RF), decision tree, and logistic regression (LR) algorithm to detect bank fraud for a synthetic dataset having 1,00,000 rows and 32 columns. Adequate preprocessing, decoding, rigorous feature engineering, validation, performance evaluation, and explanation have allowed the readers to understand the whole study. The algorithms' accuracy is similar for label encoding, which is not prescribed. Still, a significant AUC of 98% has been achieved in Gradient Boosting Models. Further application of this study can be done in real-life cases of banks, insurance, and finance institutions.