A COGNITIVE-DRIVEN MACHINE LEARNING FRAMEWORK FOR REAL-TIME FRAUD DETECTION IN ONLINE TRANSACTIONS

Abstract

Online transaction fraud is rare, adversarial and temporally structured, which makes it difficult for static classifiers and for deep models that consume every available attribute. This paper proposes a cognitive-driven machine learning framework that mimics how a human fraud analyst works: first screen the evidence that matters, then attend to the relationships among pieces of evidence, then remember the sequence of events, and finally decide. The framework uses a single mode of feature selection, Improved Information Gain (IIG), which extends classical information gain with class-imbalance-weighted entropy, split-information normalisation and a redundancy penalty. The selected features are passed to EfficientGRU-TransformerNet, which couples a linear-complexity (efficient) Transformer encoder, used to model inter-feature and inter-transaction dependencies, with a Gated Recurrent Unit (GRU) layer, used to summarise temporal behaviour compactly. A calibrated classification head with a validation-selected decision threshold separates fraudulent from legitimate transactions. The paper formalises the architecture, supplies three algorithms (IIG selection, EfficientGRU-TransformerNet forward pass, and leakage-free training and inference), analyses computational complexity, and defines a complete evaluation protocol on two public benchmarks with precision, recall, F1-score, PR-AUC, ROC-AUC and MCC as primary metrics. The proposed design targets high recall at controlled false-alarm rates and reduced model cost compared with quadratic-attention Transformers.