Abstract
This article investigates fraud detection in financial transaction networks using machine learning and graph-based typologies. The object of the study is financial transaction data, analyzed to improve the accuracy and efficiency of identifying fraudulent activities. The problem addressed is the limited generalizability and low recall of traditional fraud detection models in complex, real-world settings. To address this, a hybrid framework was developed that integrates Random Forests, neural networks, and graph-based typology indicators. Seven laundering typologies were extracted from a transaction graph – fan-in, fan-out, scatter-gather, gather-scatter, cycle, bipartite, and stacked bipartite – and used as additional features for classification. SMOTE was applied to correct class imbalance during training. Experimental results show that adding typology features significantly improves model performance. The best results were obtained with Random Forest: 98.5% accuracy, 79.1% precision, 56.3% recall, and an F1-score of 65.7%. Adding typology-based flags raised recall by 9–11 percentage points compared to models without them. Graph patterns like fan-in and fan-out were detected in 3.5–5.1% of transactions, while more complex structures such as cycle and scatter-gather appeared less frequently but correlated more strongly with known fraud. Unsupervised methods also showed promise: an autoencoder captured 60% of fraud cases among the top 2% anomalous transactions, while K-means identified 55% of fraud within flagged clusters. These methods proved useful for identifying emerging fraud types not yet labeled in training data. The model is suitable for integration into financial security systems with minimal input requirements – account IDs, timestamps, and transaction amounts—alongside basic graph analytics. Its robustness across datasets suggests strong applicability across diverse financial institutions