高维小样本下多源财税数据融合的税务风险识别模型研究

Research on a Tax Risk Identification Model Based on Multi-Source Fiscal and Tax Data Fusion in High-Dimensional Small-Sample Scenarios

摘要: 针对涉税数据多源异构、高维稀疏、风险样本稀缺及类别不均衡等问题，本文构建多源财税数据融合的税务风险识别模型。模型分别采用TabNet、轻量化LSTM和图神经网络提取结构化财税、月度经营时序和企业关联拓扑特征，并通过多头注意力机制完成特征融合；同时结合正则化、Dropout、早停和半监督伪标签扩充，以缓解小样本条件下的过拟合。基于匿名化企业涉税样本的实验结果显示，模型AUC达到0.942，F1值较主要基线模型提高12%~25%，说明其在风险样本稀缺场景下具有较好的识别效果和稳定性。研究可为税务风险筛查和企业合规自查提供参考。

Abstract: To address multi-source heterogeneity, high-dimensional sparsity, scarce risk samples, and class imbalance in tax-related data, this paper develops a tax risk identification model based on multi-source fiscal and tax data fusion. TabNet, a lightweight LSTM, and a graph neural network are used to extract structured fiscal and tax features, monthly operational time-series features, and enterprise association topology features, respectively, while a multi-head attention mechanism is introduced for feature fusion. Regularization, Dropout, early stopping, and semi-supervised pseudo-label augmentation are further adopted to alleviate overfitting under small-sample conditions. Experiments on anonymized enterprise tax data show that the model achieves an AUC of 0.942 and improves the F1-score by 12% to 25% over major baseline models, indicating good identification performance and stability when risk samples are scarce. The study provides a reference for tax risk screening and enterprise compliance self-inspection.