Detecting fraudulent activities in the banking industry is a challenging task.This article presents a novel hybrid approach that combines a rule-based system and an ensemble classifier to help and automate gambling transaction detection.
The rule-based system generates labelled data by applying predefined rules to the transactions, and the tree-based ensemble classifier uses the labelled data to learn and differentiate between gambling and non-gambling transactions. Three types of tree-based ensemble classifiers, namely Random Forest, XGBoost, and LightGBM, were selected and compared. The advantage of using tree-based classifiers is that they not only allow for accurate predictions but also facilitate the extraction of rules that could be used to update the rule-based system. The Our proposed hybrid approach operates on the principle that both the rule-based system and the machine learning classifier complement each other in a validation cycle, and this combination improves the accuracy of the results and produces a robust model. In addition, to increasing the labelled data that makes the machine learning classifier more reliable, the ensemble machine learning classifier also adds new desirablecan help detect items transactions to the results that were not detected by the rule-based system and as they have different and new behaviors compared to this of new types of this kind of fraud.
The results show that LightGBM achieved the best performance with an accuracy of about 97% and an F1-measure of about 96%. The results also demonstrate that the proposed hybrid approach is effective and feasible in illegal gambling detection.