Developing an Unsupervised Feature Extraction Framework for Financial Transaction Fraud Detection using Autoencoders and Temporal Logic

Fraud detection in financial transactions is challenged by extreme class imbalance and a lack of labeled data. In this work, we introduce a feature-extraction framework for the problem of mobile-money fraud detection on the PaySim synthetic transaction dataset (~6.3M, 0.13% fraud) that combines a deep autoencoder with a temporal, rule-based logic component. The autoencoder is trained solely on negative samples and learns a low-dimensional latent embedding whose reconstruction error is used as an unsupervised anomaly signal. This is then combined, via a rank-normalized additive fusion, with a temporal logic scoring function that encodes a set of behavioral and balance-consistency rules relating to account draining and mule-account activity. An ablation study of fusion strategies and mixing weights is used to determine an optimal configuration for this setup. On a held-out test set of the true fraud rate, the resulting framework achieves PR-AUC 0.336 and ROC-AUC 0.992, outperforming the autoencoder alone (PR-AUC 0.180) as well as standard unsupervised baselines (Isolation Forest, Local Outlier Factor, One-Class SVM). These results demonstrate that a learned latent anomaly signal can be meaningfully combined with temporal rule-based logic to greatly improve the precision-recall performance of mobile-money fraud detection in highly imbalanced settings.