Random Forest Deep Ensemble for Tax Audit Case Selection
Ke Yang, Xu Chen · Applied Sciences · 2026
Tax is the principal source of a nationś fiscal revenue. Countless tax fraud behaviors have been reported through a wide variety of techniques. This paper proposes a three-stage Random Forest Deep Ensemble algorithm for tax audit case selection. The framework combines class-based limited undersampling, Random Forest base learners, and a Deep Belief Network ensemble. The transition from the Random Forest stage to the DBN stage is explicitly modeled as a stacking process: the scalar outputs of the K CART base classifiers are concatenated into a K-dimensional meta-feature vector and then learned by the DBN meta-learner. The study further introduces the TXFCN dataset, a cost–benefit-oriented metric named RCITA, cross-validation robustness checks, feature-importance interpretation, and limitations related to sample selection bias and external validity. The empirical evaluation on a tax fraud dataset shows that the proposed method has significant advantages in detecting potential tax fraud.