Hybrid Graph Neural Networks and Transformer Models for Regulatory Data Lineage in Banking

Regulatory compliance in the banking sector increasingly depends on transparent and auditable data lineage across complex financial systems. Traditional rule-based and relational approaches often fail to capture the dynamic, heterogeneous, and high-dimensional nature of regulatory data flows. This study introduces a hybrid framework that integrates Graph Neural Networks (GNNs) with Transformer models to address these challenges. GNNs are leveraged to model relational structures within financial data ecosystems, capturing entity-level dependencies and transaction-level connectivity, while Transformers provide contextual representation learning for sequential and textual regulatory metadata. The hybrid architecture enables scalable lineage tracing, anomaly detection, and regulatory reporting with higher precision and explainability. Experimental results on synthetic and real-world banking datasets demonstrate significant improvements in lineage accuracy, compliance monitoring efficiency, and interpretability compared to baseline methods. The findings highlight the potential of combining graph-based learning with attention-driven architectures to meet evolving regulatory requirements, enhance data governance, and strengthen trust in financial decision-making processes.