Tackling fraud detection with an enhanced Kepler optimization and ghost opposition-based learning

TL;DR

This paper introduces a new fraud detection method combining enhanced optimization algorithms and sampling techniques to improve accuracy and efficiency in detecting fraud and malware.

Contribution

The novel BKOA-GOBL method integrates Binary Kepler Optimization with Ghost Opposition-Based Learning for improved feature selection and fraud detection.

Findings

BKOAG-GBL achieved up to 99.96% accuracy on some datasets with significant feature reduction.

The method outperformed 12 other algorithms in accuracy and computational efficiency.

Lower performance on the Real vs Fake Job Postings dataset highlights detection challenges in complex cases.

Abstract

The growing prevalence of fraud and malware, fueled by increased online activity and digital transactions, has exposed the shortcomings of conventional detection systems, particularly in handling novel or obfuscated threats, class imbalance, and high-dimensional data with many irrelevant features. This underscores the need for robust and adaptive detection methodologies.

This study proposes an advanced Fraud Detection (FD) methodology, BKOA-GOBL, that enhances the Binary Kepler Optimization Algorithm (BKOA) by integrating Ghost Opposition-Based Learning (GOBL) to improve Feature Selection (FS). The BKOA dynamically models gravitational attraction, planetary motion mechanics, and cyclic control to maintain a balance between exploration and exploitation. At the same time, the GOBL enhances broader search diversification and prevents early convergence, allowing the local optimum to be avoided. The Random Under-Sampling (RUS) technique is utilized to mitigate the class imbalance in fraud benchmarks.

Experimental validation is conducted on five real-world benchmarks, including the Australian, European, CIC-MalMem-2022, Synthetic Financial Transaction Log, and Real vs Fake Job Postings datasets, using k-Nearest Neighbors (K-NN) and XGBoost (Xgb-tree) classifiers. The BKOA-GOBL achieves outstanding performance, reaching classification accuracies up to 99.96% in some benchmarks and corresponding feature reduction rates up to 81.82%. Precision, recall, ROC_AUC, and F1-scores were consistently high across most benchmarks, demonstrating reliable and balanced detection. However, some challenging benchmarks—such as the Real vs Fake Job Postings dataset using k-NN classifier—returned lower scores (Precision = 76.14%, Recall = 66.55%, F1-score = 71.00%, and ROC_AUC = 74.15%), reflecting the difficulty of the problem. Comparative analyses against 12 recent Metaheuristic Algorithms (MHAs) and Machine Learning (ML) classifiers confirmed BKOA-GOBL's dominance in terms of accuracy and computational efficiency. Its statistical superiority is confirmed by the Wilcoxon rank-sum test, underscoring its robustness, adaptability, and effectiveness in high-dimensional fraud and malware detection tasks and real-world fraud and malware detection scenarios.

Figures