| dc.description.abstract | Tax evasion is a serious problem in all countries. If the perceived benefits of evasion outweigh the perceived costs then, if it is possible, individuals will evade taxes. Owing to the lack of enough tax officers, reported cases cannot be audit one by one. That is, the audit probability determines how many reported cases will be randomly selected and reexamined. Therefore, high audit probability and high penalty rate usually may lead to a decrease of evasion. A crucial problem to be encountered is that a high audit probability needs high resources. Therefore, how to develop a computerized detection method becomes a very demanding challenge. In this thesis, a neural-network-based method is proposed to implement an efficient detection system. First, four different feature extractors; such as Pearson’s correlation, Principal component analysis, Fisher ratio, and Statistics quantity method are employed to extract effective features from reported cases. Then a Hyper Rectangular Composite Neural Networks(HRCNNs),multi-layer perceptrons(MLPs) with the backpropagation algorithm, learning vector quantization(LVQ)networks, binary logistic regression method, and decision Tree are employed to implement detection systems. The tax data sets were collected from National Tax Administration Northern Province Ministry of Finance in 1998 and 1999, respectively, the data sets were splitted into a training data set consisting of reported cases in 1998 and a testing data set consisting of reported cases in 1999. Simulation results show that the computerized detection systems outperformed the present auditing process. Therefore, these detection systems provide an alternative tool for preventing tax evasion. | en_US |