FinStack-Net: Hierarchical Feature Crossing and Stacked Ensemble Learning for Financial Fraud Detection

Detecting gambling and fraudulent accounts from transaction records and static account data is hard. Existing methods often cannot show complex feature interactions well, and they struggle when data are imbalanced or high-dimensional. This paper presents FinStack-Net, a hierarchical ensemble learning framework that uses layered feature construction, different base models, and meta-learning optimization. FinStack-Net has a hierarchical cross-feature module that creates first- and second-order interactions and uses mutual information screening and Lasso pruning to keep useful features. It then stacks gradient-boosting trees (LightGBM and CatBoost) and a deep neural network with residual connections and attention to capture both structural and nonlinear patterns. A meta-learning layer fuses these model outputs with logistic regression, and Optuna is used to run a Bayesian search for all hyperparameters. Tests on real data show that FinStack-Net outperforms single models and other ensembles, making it practical for real-world financial risk control.