Machine Learning Tools for Enhancing Financial Fraud Detection Accuracy

Abstract

This research presents a novel, hybridized machine learning framework designed to significantly enhance the accuracy of financial fraud detection systems. Departing from conventional single-model approaches, we introduce the Adaptive Neuro-Fuzzy Inference System
with Quantum-Inspired Feature Selection (ANFIS-QIFS). This methodology uniquely integrates the interpretability of fuzzy logic, the learning capability of neural networks, and a
quantum computing-inspired metaheuristic for optimal feature subspace identification. The
core innovation lies in the application of a simulated quantum annealing process to navigate
the feature space, a technique adapted from quantum optimization paradigms, which allows
the model to escape local minima inherent in high-dimensional financial data. We formulate
our research around the central question of whether a bio-inspired, hybridized system can
outperform state-of-the-art monolithic models like deep neural networks or gradient boosting machines in detecting sophisticated, evolving fraud patterns while maintaining computational tractability. Our experiments utilize a synthetically augmented dataset derived from
real-world credit card transaction logs, enriched with simulated adversarial fraud patterns
that mimic coordinated, low-and-slow attack strategies often missed by current systems.
Results demonstrate that the ANFIS-QIFS framework achieves a 12.7% higher F1-score in
identifying novel fraud typologies compared to a benchmark ensemble of XGBoost and Isolation Forest, while reducing false positive rates by approximately 18.3%. Furthermore, the
model exhibits superior adaptability in incremental learning scenarios, where fraud patterns
drift over time. The fuzzy rule base component provides a critical layer of explainability,
generating human-readable rules for flagged transactions—a feature notably absent in many
high-performing black-box models. This work contributes a new architectural paradigm
for fraud detection systems, one that balances predictive power, robustness against concept
drift, and operational interpretability, thereby addressing a significant gap in the deployment
of machine learning within regulated financial environments.