A META-BOOSTING ENSEMBLE MODEL WITH ENHANCED FEATURE GROUPING FOR ROBUST FRAUD DETECTION

Despite the complexity of conducting fraud detection with ever changing types of fraudulent activities, fraud detection in financial transactions is still a challenge. Based on these problems, this research then proposes a novel meta-boosting framework for fraud detection with feature grouping, parallel boosting models and meta-features generation. The third group features from this dataset are used to extract distinct fraud patterns, that is, features are divided into three categories—card, device and transaction features. Three gradient boosting models, LightGBM, XGBoost, and CatBoost, are trained in parallel, and their results are compiled by ta Meta Boosting model which makes use of a meta feature construction to make prediction true. We also introduce a Custom Fraud Detection Index (FDI), a collective of evaluation metrics that combine to form one score for overall evaluation. Results of the proposed approach are shown to have an FDI of 0.7091, AUC of 0.9721, Log-Loss of 0.9007, and MCC of 0.5132, and outperform the performance of individual models. The meta boosting strategy helps a lot to improve fraud detection, especially when dealing with imbalanced datasets with high Precision Recall AUC (0.9007). In contrast to the conventional single model approach, our solution is able to effectively combine multiple feature sets to create a more robust, generalizable and scalable fraud detection system. It shows that in large scale financial application; meta boosting can increase fraud detection accuracy.