Cross-regional enterprise tax evasion identification method based on PU learning

Abstract

The invention discloses a cross-regional enterprise tax evasion identification method based on PU learning, and the method comprises the steps: firstly generating a label of a label-free sample in a source domain through a small number of positive samples with labels in the source domain based on PU learning; secondly, constructing a cross-regional enterprise tax evasion identification model through a domain adaptation method of maximizing classifier difference; and finally, performing tax evasion identification on the tax payment data of the target domain by using the trained tax evasion identification model. Under the condition that source domain tax payment data only has positive samples and a large number of unmarked samples, the purpose of establishing a tax evasion identification model for a target domain without a label of the tax payment data is achieved.

Description

technical field [0001] The invention belongs to the technical field of tax inspection, and in particular relates to a method for identifying tax evasion by cross-regional enterprises based on PU learning. Background technique [0002] In recent years, with the rapid development of computer science and Internet technology, the combination of domestic tax collection management and information technology has become a development trend in the era of big data. Among them, with the development of the "Golden Tax Phase III" and "Internet + Taxation", the domestic taxation department has accumulated a large amount of tax-related data. How to make full and effective use of these tax-related data to implement tax audit has become a new challenge. [0003] Due to the limitation of manpower and time, the tax inspection department cannot inspect every taxpayer. Generally, taxpayers who are suspected of tax evasion are sampled through audit selection, and the audit is carried out on thes...

AI Technical Summary

Problems solved by technology

However, in the field of taxation, the case selection method based on data mining still faces the following problems: due to the uneven economic development among provinces and cities, there are great differences in the scale and characteristics of taxation data; However, some areas (target domains) lack or even have no marked data; because tax audits are limited by limited manpower and material resources, even in developed areas, it is impossible to audit all enterprises. In practice, tax audits are often used The sampling inspection method makes the areas with sufficient labeled data (source domain) only have a small number of marked positive labels, so it is impossible to establish a unified tax evasion identification model

[0008] Document 2 proposes a corporate tax evasion identification method based on deep adversarial transfer learning. By using the tax data in the source domain, a tax evasion identification model suitable for the target domain is constructed to solve the problem that the target domain cannot be detected due to the lack of labeled data. Problems of identification and modeling of corporate tax evasion

[0009] The method described in the above literature mainly has the following problems: the identification model of literature 1 requires the characteristics of tax data to be independent and identically distributed and needs to rely on certain expert knowledge, which cannot solve the problem of different distribution of tax data characteristics in different regions and cannot effectively identify tax evasion across regions this question

The method in Document 2 requires that the source domain tax data have all labels, but in actual tax scenarios, the source domain tax data often only has positive samples and a large number of unlabeled samples. The premise of this method does not match the actual tax scenario. Labeling takes a lot of time and cost, making the model less viable

Images