DF-NFLF-TO: A Deep Fusion Neuro-Fuzzy–Logistic Model for Fraud Detection under Extreme Class Imbalance

All over the world people suffer from financial fraud in credit cards so need robust model to detect the fraudulent transaction over the large amount of records. There is a problem of highly imbalanced transaction data because very small number of fraudulent records in datasets. Conventional deep learning models report high accuracy under such conditions but fail to reliably identify minority fraud instances, and lead to poor recall and limited practical utility. To address this challenge, this paper proposes Deep Fusion Neuro-Fuzzy–Logistic model with validation-driven threshold optimization for robust fraud detection under extreme class imbalance. The proposed model combines the probabilistic logistic learning, neuro-fuzzy inference, and deep representation learning into a unified late-fusion model to enable the complementary decision evidence from heterogeneous models. Unlike fixed decision thresholds commonly adopted in prior studies, DF-NFLF-TO used to validation-based threshold optimization strategy to explicitly control false positive rates while preserving meaningful fraud capture. The experimental model conducted on a large-scale financial transaction dataset shows that the proposed model significantly improves imbalance-aware performance with the accuracy score of 99.88%. The final outcome shows that combined deep fusion with threshold optimization gives more trustworthy solution for actual fraud detection.