Developing a Model to Identify the Unreal Returns on Value Added Tax Using Data Mining Approach

The tax evasion is a constant concern for the tax administrations, especially in developing countries. Due to the large number of Value Added Tax (VAT) returns and resource constraints or their unaffordable investigation, it is necessary to develop a mechanism to identify dishonest taxpayers on the basis of historical data in large databases in this area. In this research via a survey approach, eighteen variables that potentially affecting the identification of unreal statements are identified and using some data provided from VAT returns and performance, their impact on the detection of tax fraud are investigated. After preprocessing of the data based on filtering techniques, ten influential factors in predicting the tax records are set. Genetic Algorithm is reduced the potential independent variables to seven influential variables. The variable for the status of the tax records in terms of fraud is defined and to predict their situation, the prediction model with a decision tree approach, which is a data mining method, is developed. Implementations based on decision tree and ensemble methods of Bagging and Boosting on observations indicate that the decision tree and ensemble Bagging and Boosting methods which using ten predictive factors, have the ability to predict the status of the records with the accuracy of 82.14 percent. A set of rule in order to preprocess the record is identified that can identify potential fraud before it is reviewed by the tax auditors.