Authors: Sukhpreet Kaur, Assistant Professor Ms. Jaspreet Kaur
Abstract: Credit card use has expanded rapidly, and with it the opportunities for fraud that is hard to distinguish from legitimate spending. Machine learning has become the main research response, yet published results are difficult to compare and harder to translate into operational practice. This review synthesises 22 scholarly publications drawn from the reference base of a master’s dissertation on the topic and reorganizes them into thematic strands: classifier benchmarking, feature engineering and hybrid designs, sequence and generative approaches, data availability and class imbalance, and evaluation, explainability and concept drift. The method is a reconstructed, non-systematic thematic synthesis; no new database search or screening was carried out. The reviewed work indicates that tree-based ensembles are the most frequently applied and often the strongest supervised models, that behaviorally informed features and hybrid designs add value beyond the choice of algorithm, and that accuracy is an inadequate measure when fraud makes up a small fraction of one per cent of transactions. The evidence base, however, rests heavily on one anonymized public dataset and on confidential industrial data, which limits reproducibility and generalization. Five gaps stand out: scarce real data, non-comparable evaluation protocols, concept drift and delayed labels, limited explainability, and the absence of realtime and local validation, including for the Indian market. Research directions are proposed for each gap, and standardized temporal evaluation is argued to be the most immediate need.
DOI: