Fraud Detection in Banking Data by Machine Learning Techniques

With the rise of technology and e-commerce, credit card transactions have surged, leading to an increase in banking fraud and associated costs. This study explores detecting fraudulent transactions using class weight-tuning hyperparameters, optimized through Bayesian methods, to handle un balanced data. We suggest using weight adjustment as a solution preprocessing step and evaluate XGBoost and LightGBM through a majority voting ensemble learning method. Experiments on real-world data show that LightGBM and XGBoost achieve high performance, with ROC-AUC of 0.95; precision 0.79, recall 0.80 and F1 score is 0.79. Incorporating deep learning for hyper parameter tuning further enhances performance, achieving ROC-AUC of 0.94, precision 0.80 and recall 0.82, and F1 score 0.81, surpassing current methods.