A Model for Identification Tax Fraud Based on Improved ID3 Decision Tree Algorithm andMultilayer Perceptron Neural Network

Tax revenues are one of the most important sources of governments and cover a large portion of government spending. In recent years, fraud in financial statements and tax returns has increasingly become a serious problem for businesses, governments and investors. Most taxpayers are looking for a way to manipulate their financial statements and reduce their taxable profits. Therefore, identifying tax fraudsters and companies that cheat on financial statements has become a vital issue for the government.The purpose of this study is to present a model that uses the improved ID 3decision tree algorithm. Also, to improve its performance and accuracy, it was combined with multilayer perceptron neural networks optimized by genetic algorithm to select financial ratios associated with tax fraud and reduce computational overhead. The tree in the proposed model has the lowest depth possible, so it has high velocity and low computational overhead. For this purpose, the financial statements of 06companies listed in Tehran Stock Exchange during - 4330 4331were studied and 41financial ratios were extracted. By ANOVA test, 33ratios and finally by neural networks 7ratios related to tax fraud was selected as the model input data. The proposed model, with %4411accuracy, has been successful in identifying fraudulent companies with the highest accuracy and predictive power over the adaboost algorithms.