Can artificial intelligence detect fraudulent financial transactions accurately when confirmed fraud labels are scarce?
This study introduces SAE-IF, a hybrid semi-supervised framework that combines a sparse autoencoder with an isolation forest to detect unusual patterns in credit-card and digital-payment transactions.
The sparse autoencoder learns compressed representations of transaction behaviour without using class labels in its reconstruction objective. The isolation forest then examines these learned representations and assigns anomaly scores to transactions that differ from normal patterns.
To improve reliability, the researchers used a leakage-aware experimental design with separate training, validation, and testing sets. Data preprocessing parameters were calculated from the training data only, while Bayesian optimisation with Optuna was used to select the model settings, fusion weight, and decision threshold.
The framework was tested on three substantially different datasets: the ULB/Kaggle credit-card dataset, the IEEE-CIS Fraud Detection dataset, and the PaySim mobile-money dataset.
On the ULB/Kaggle dataset, SAE-IF achieved a precisionโrecall area under the curve of 0.832, an F1-score of 0.812, precision of 0.915, and recall of 0.728. It reduced false positives to 42, outperforming the standalone sparse autoencoder and isolation forest models.
The framework also recorded PR-AUC values of 0.801 on IEEE-CIS and 0.845 on PaySim, while maintaining precision above 0.90 across all three datasets.
A weighted fusion mechanism gave greater importance to the isolation forest anomaly score while retaining complementary information from the autoencoderโs reconstruction error. This combination produced more reliable fraud detection than either model used independently.
The model processed transactions in approximately 1.4โ1.6 milliseconds on standard CPU hardware under controlled offline testing conditions, indicating potential for near-real-time financial monitoring.
Visual analysis of the learned feature space showed that legitimate transactions formed a concentrated cluster, while fraudulent transactions appeared mainly at the boundaries or in separate regions.
Although fully supervised methods achieved higher overall performance, SAE-IF outperformed the unsupervised alternatives examined and offers an important advantage in environments where fraud labels are limited, delayed, or unreliable.
The findings demonstrate that combining deep representation learning with anomaly isolation can strengthen fraud screening, reduce false alerts, and support safer online banking, mobile payments, e-commerce, and other digital financial services.
๐ Read the full article here:
https://doi.org/10.46481/jnsps.2026.3401
Published in: Journal of the Nigerian Society of Physical Sciences