Investigating the Usefulness of Relief Selection Variable Method in Improving Tax Evasion Prediction Outcomes Using Data Mining

The present study examines the usefulness of  Relief method and data mining in predicting tax evasion of listed companies in Tehran Stock Exchange (TSE) using accounting data and decision tree patterns in two situations: with and without the phase of selecting variables. The statistical population of this study includes all companies accepted in TSE from 2005 to 2015, and the research sample included 1.081 company-years. One-way ANOVA, independent sample t-test, decision tree algorithms, and the Relief method of selecting variables were used for data analysis. Data was analyzed using SPSS and Weka softwares. The results of  Relief algorithm showed that ratio of operating profit to total assets, ratio of return on assets, and market value of company are more appropriate variables than other variables for predicting tax evasion. In addition, the results of one-way ANOVA showed that the difference in prediction accuracy of different decision tree methods is statistically significant. However, when each of these algorithms compared with other separately, both states of with and without the phase of selecting optimal variables, the results showed that only LMT algorithms results were significantly different with each other. In other algorithms, even though the results improved but this was not statistically significant. In other words, using relief method does not improve the results in all algorithms.