An intelligent credit card fraud detection approach based on semantic fusion of two classifiers

• Published in: Soft computing (Berlin, Germany) Germany), 2020, Vol.24 (2), p.1243-1253
• Main Author: Darwish, Saad M.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Classification; Classifiers; Clustering; Clusters; Communications systems; Computational Intelligence; Control; Credit card fraud; Cybercrime; Data mining; Datasets; Electronic banking; Engineering; False alarms; Foraging behavior; Fraud; Fraud prevention; Geographical locations; Literature reviews; Machine learning; Mathematical Logic and Foundations; Mechatronics; Methodologies and Application; Neural networks; Online banking; Robotics; Search algorithms; Semantics; Statistical methods; Swarm intelligence; User behavior
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-019-03958-9

The increased usage of credit cards for online and regular purchases in E-banking communication systems is vulnerable to credit card fraud. Data imbalance also poses a huge challenge in the fraud detection process. The efficiency of the current fraud detection system (FDS) is in question only because they detect the fraudulent activity after the suspicious transaction is done. This paper proposes an intelligent two-level credit card fraud detection model from highly imbalanced datasets, relying on the semantic fusion of k -means and artificial bee colony algorithm (ABC) to enhance the classification accuracy and speed up detection convergence. ABC as a second classification level performs a kind of neighborhood search combined with the global search to handle the inability the k -means classifier to discover the real cluster if the same data is inputted in a different order it may produce different cluster. Besides, the k -means classifier may be surrounded by the local optimum as it is sensitive to the initial condition. The advised system filters the dataset’ features using a built-in rule engine to analyze whether the transaction is genuine or fraudulent based on many customer behavior (profile) parameters such geographical locations, usage frequency, and book balance. Experimental results indicate that the proposed model can enhance the classification accuracy against the risk coming from suspicious transactions, and gives higher accuracy compared to traditional methods.