Tax evasion, an unlawful practice in which taxpayers deliberately conceal information to avoid paying tax liabilities, poses significant challenges for tax authorities. Effective tax evasion detection is critical for assisting tax authorities in mitigating tax revenue loss. Recently, machine-learning-based methods, particularly those employing positive and unlabeled (PU) learning, have been adopted for tax evasion detection, achieving notable success. However, these methods exhibit two major practical limitations. First, their success heavily relies on the strong assumption that the label frequency (the fraction of identified taxpayers among tax evaders) is known in advance. Second, although some methods attempt to estimate label frequency using approaches like Mixture Proportion Estimation (MPE) without making any assumptions, they subsequently construct a classifier based on the error-prone label frequency obtained from the previous estimation. This two-stage approach may not be optimal, as it neglects error accumulation in classifier training resulting from the estimation bias in the first stage. To address these limitations, we propose a novel PU learning-based tax evasion detection framework called RR-PU, which can revise the bias in a two-stage synergistic manner. Specifically, RR-PU refines the label frequency initialization by leveraging a regrouping technique to fortify the MPE perspective. Subsequently, we integrate a trainable slack variable to fine-tune the initial label frequency, concurrently optimizing this variable and the classifier to eliminate latent bias in the initial stage. Experimental results on three real-world tax datasets demonstrate that RR-PU outperforms state-of-the-art methods in tax evasion detection tasks.

In label-noise learning, accurately identifying the transition matrix is crucial for developing statistically consistent classifiers. This task is complicated by instance-dependent noise, which introduces identifiability challenges in the absence of stringent assumptions. Existing methods use neural networks to estimate the transition matrix by initially extracting confident clean instances. However, this extraction process is hindered by severe inter-class imbalance and a bias toward selecting unambiguous intra-class instances, leading to a distorted understanding of noise patterns. To tackle these challenges, our paper introduces a Class Rebalance and Geometric Regularization-based Framework (CRGR). CRGR employs a smoothed, noise-tolerant reweighting mechanism to equilibrate inter-class representation, thereby mitigating the risk of model overfitting to dominant classes. Additionally, recognizing that instances with similar characteristics often exhibit parallel noise patterns, we propose that the transition matrix should mirror the similarity of the feature space. This insight promotes the inclusion of ambiguous instances in training, serving as a form of geometric regularization. Such a strategy enhances the model's ability to navigate diverse noise patterns and strengthens its generalization capabilities. By addressing both inter-class and intra-class biases, CRGR offers a more balanced and robust classification model. Extensive experiments on both synthetic and real-world datasets demonstrate CRGR's superiority over existing state-of-the-art methods, significantly boosting classification accuracy and showcasing its effectiveness in handling instance-dependent noise.