Tax authorities worldwide face a significant challenge: identifying tax evaders within a massive pool of taxpayers where only a small fraction are confirmed offenders. This "needle in a haystack" scenario is a classic example of Positive and Unlabeled (PU) learning. In this setting, we have a set of confirmed tax evaders (Positive) and a vast majority of taxpayers who are Unlabeled—meaning they could be law-abiding citizens or evaders who haven't been caught yet.
Traditional PU learning algorithms often struggle because they require an accurate prior estimate of the label frequency (the proportion of total evaders that have been identified). If this initial estimate is incorrect, the subsequent classification performance suffers. This paper, presented at AAAI-24, introduces RR-PU, a robust framework designed to overcome these limitations through a synergistic two-stage approach.
The RR-PU Framework
RR-PU addresses the "catch-22" of PU learning by treating the label frequency not as a fixed constant, but as a learnable parameter that can be refined during the training process. The framework consists of two primary stages:
- Label Frequency Initialization (Regrouping-MPE): The model first uses Mixture Proportion Estimation (MPE) to generate an initial guess of the label frequency. To improve accuracy, the authors implement a "regrouping" technique. This helps the algorithm satisfy the "irreducible assumption," ensuring that the distributions of "good" and "bad" taxpayers are sufficiently distinct for the model to differentiate between them.
- Revision Stage (The "Synergy"): This is the core innovation of RR-PU. Instead of relying solely on the initial estimate, the framework introduces a trainable slack variable ($\Delta T$). As the backbone neural network learns to classify taxpayers, it simultaneously adjusts this slack variable to fine-tune the label frequency estimate.
By optimizing a weighted risk function, the classifier and the frequency estimator improve each other concurrently. This "closing of the loop" allows the model to correct latent biases from the first stage, leading to more robust detection.
Experimental Evaluation
The researchers evaluated RR-PU using three real-world datasets provided by Chinese tax bureaus: TaxS, TaxH, and TaxZ. These datasets include complex features such as value-added invoices, tax amounts, debt, and profit ratios. To simulate the PU learning environment, the researchers took fully labeled data and "hid" 50% of the positive labels, treating them as unlabeled.
To handle the complex transaction data, a graph-based method (PnCGCN) was employed to extract features from invoices, ensuring the model had a rich representation of taxpayer behavior.
Key Results
RR-PU consistently outperformed a wide range of baselines, including specialized tax detection models like Eagle and FBNE-PU, as well as general PU learning algorithms like nnPU and Dist-PU.
As shown in the results, RR-PU achieved the highest Accuracy, F1-score, and AUC across all three datasets. On the TaxS dataset—noted as the most challenging—RR-PU improved accuracy by approximately 3.5% over the next best method. Furthermore, ROC curve analysis confirmed that RR-PU maintains a higher True Positive Rate at lower False Positive Rates, a critical requirement for tax authorities who must manage limited audit resources efficiently.
Ablation studies further confirmed that both the "Regrouping" and "Revision" stages are essential; removing either component resulted in a significant drop in performance, validating the synergistic design of the framework.
Conclusion and Future Work
RR-PU represents a significant advancement in tax evasion detection by making the label frequency estimate a learnable, self-correcting part of the model. While the framework is highly effective, the authors note that future work could focus on improving the interpretability of the model—helping auditors understand why a specific taxpayer was flagged—and optimizing the MPE component for even higher-dimensional datasets.
Summary updated