TITLE:
SHAP-Driven Interpretability in Financial Fraud Detection: A Multimodal Data Approach
AUTHORS:
Nie Hui
KEYWORDS:
Financial Fraud Detection, Multimodal Data Fusion, Explainable AI, Textual Semantic Analysis, Ensemble Learning
JOURNAL NAME:
Journal of Computer and Communications,
Vol.13 No.12,
December
19,
2025
ABSTRACT: Improved accuracy in predicting corporate financial fraud significantly enhances regulatory efficiency and market stability. However, detecting increasingly sophisticated fraud patterns remains challenging due to data heterogeneity and model opacity. This study proposes an innovative, multimodal framework that integrates corporate financial indicators, organizational structures, and semantic features from annual reports. We employ BERT-based semantic extraction on the Management Discussion & Analysis (MD&A) section of an annual report, reduce dimensionality via PCA, and fuse features with corporate financial/organization structural metrics. Ensemble tree models (CatBoost/XGBoost/LightGBM) are optimized for fraud prediction, while SHAP values quantify the contributions of individual features. Experimental results demonstrate a peak ROC-AUC of 0.859, with key findings revealing that: 1) Significant asset transactions, future outlook, and operational overview are the most predictive MD&A contents; 2) Financial indicators dominate feature importance (50% of top predictors); 3) Annual report similarity and tone serve as critical textual red flags. This framework provides regulators with actionable insights through model interpretability, thereby advancing early-warning systems for financial misconduct.