A Generative and Cluster-Aware Multi-model Architecture for Detecting Heterogeneous Fraud in Imbalanced Transaction Data
Workshops of 26th International Conference on Computational Science and its Applications, ICCSA 2026, Braga, Portekiz, 30 Haziran - 03 Temmuz 2026, cilt.16758 LNCS, ss.16-32, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 16758 LNCS
- Doi Numarası: 10.1007/978-3-032-30497-1_2
- Basıldığı Şehir: Braga
- Basıldığı Ülke: Portekiz
- Sayfa Sayıları: ss.16-32
- Anahtar Kelimeler: Class Imbalance, CTGAN, Fraud Detection, Mixture-of-Experts, UMAP
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
With the rapid development of digital banking, the issue of financial fraud detection has grown even more pressing. Real-world fraud systems have two major challenges: extreme class imbalance and behavioral variability of fraudulent activities, which are frequently not captured by monolithic global models. We propose a cluster-aware multi-model framework that simultaneously tackles both challenges. To handle the class imbalance efficiently, we first employ a Conditional Tabular Generative Adversarial Network (CTGAN) to generate high-fidelity synthetic fraud examples. Second, we cluster fraudulent transactions into behaviorally consistent subgroups using non-linear dimensionality reduction (UMAP) and density-based clustering (DBSCAN). This breakdown enables a Mixture-of-Experts (MoE) architecture in which specialized classifiers are trained for different fraud behaviors. We evaluate our approach on a popular benchmark for credit card fraud and show that it outperforms classical resampling methods and strong global baselines. Ultimately, the suggested technique detects infrequent and confined fraud patterns more effectively, resulting in a much higher F1-score and operational viability.