Network-Enhanced Neuro-Symbolic Framework for Financial Fraud Detection: Integrating Corporate Network Topology with Interpretable AI

Corporate fraud imposes substantial costs on investors and the broader economy, yet existing detection methods face a fundamental trade-off between accuracy and interpretability. Traditional rule-based approaches provide transparency but achieve limited performance (F1-scores 0.55-0.65), while machine learning models demonstrate superior accuracy (87-94%) but lack regulatory-required interpretability. We address this challenge by proposing a network-enhanced neuro-symbolic framework that integrates corporate network topology with interpretable AI. Using 15,724 restatement cases from 9,951 U.S. companies (2000-2024), we construct a board interlock network from 561,306 director records and extract network features that capture fraud contagion patterns. Our framework combines LSTM neural networks for pattern recognition, domain-specific symbolic rules for transparency, and an attention mechanism for optimal integration. Results demonstrate that network features contribute 31.2\% of predictive power, with fraud neighbor ratio emerging as the strongest predictor. The integrated framework achieves superior performance (F1=0.891) compared to pure neural (F1=0.723) or symbolic (F1=0.651) approaches, while maintaining interpretability through SHAP analysis and symbolic rule activation. Temporal validation across 15 time periods confirms model robustness. Implementation shows processing times under 2.3 seconds per company with 38% reduction in audit time. These findings establish network topology as a fundamental fraud detection signal and demonstrate that neuro-symbolic integration can resolve the accuracy-interpretability trade-off in regulated environments.