運用資料探勘提昇虛設行號選案之研究

Researches on the Improvement of Fictitious Company Detection Performance using Data Mining Techniques

林明冠(Ming-Kuan Lin)

指導教授 ： 盧以詮

元智大學/資訊學院/資訊管理學系/碩士(2011年)

摘要

摘 要

營業稅是一種對提供應稅勞務、轉讓無形資產或銷售不動產的營業人根據其營業額的徵收稅。作為三大流轉稅之一，目前營業稅已成為地方稅收的主體稅種營業稅，也成為目前營業稅逃漏的主要方式，其中虛設行號問題是現階段侵蝕營業稅稅收之最主要原因。
本研究運用資料探勘技術協助稽徵業務作業可能性進行了研究。運用對預測警示問題有良好成效的支撐向量機（support vector machine，SVM）技術，建立營業稅的虛設行號警示選案模型，期望能達到事先防範逃漏營業稅。本研究使用從民國78-97年共1,324,188筆營業稅稅籍資料登記為批發零售業之營業人，並連結公司行號負責人之個人戶籍資料及其他課稅資料（遊民、殘障資料），以支撐向量機技術建構一套系統化且能辨識虛設行號特徵因數的新選案模式。以修正後之訓練資料1,576筆作為樣本，進行支撐向量機的訓練，使訓練完成的支撐向量機能對預測組資料進行預測分類。
經實證研究結果，支撐向量機比類神經網路更容易實現，能夠實現更快和更穩定的訓練；由於其嚴密的數學理論基礎，能夠達到較高的選案正確率。在變數敏感度分析，結果顯示營業人或負責人有巨額欠稅、現在擔任其他公司負責人及負責人擔任其他異常營業人之負責人係未來稅務機關查緝虛設行號案件之重要因數，可作為稅務機關在選案時的參考

關鍵字

資料探勘 ； 支撐向量機 ； 虛設行號

並列摘要

ABSTRACT

The business tax is a consumption tax charged at the point of purchase for certain goods and services. Nowadays, the business tax has become one of the important taxes for regional taxes, because the tax dodging and the tax evasion are very serious, and the amount of amercement about the business tax dodging and the business tax evasion is the highest among all the tax amercement. Getting to the bottom of the issue, the reason is that the fictitious company is the prime criminal of corrupting the business tax.
In this paper, we used an efficient data mining machine, support vector machine (SVM), which proves to perform well in prediction problems, in fictitious company detection. A fictitious company presentiment model is established to prevent the the tax dodging and the tax evasion. All the materials (including the management and the company) of business taxes cases from 1990 to 2010 is used to test our established support vector machine based prediction system. The pre-processing result of 1,576 samples is served as the training data of SVM, and the others are used to investigate the performance of SVM.
The simulation and the analysis on the database indicate that SVM is more feasible than the artificial neural network, and it has rapider convergence and more stable training result than the artificial neural network, so achieving high accuracy in prediction. By an analysis on the sensitivity on the variables, we found that the whether the principal has the substantive due tax, occupy the other (and the exceptional) companies or not are more important than the others, which can be used as a reference in fictitious company detection.

並列關鍵字

Data mining ； Support vector machine ； Fictitious Company

參考文獻

[10]. 廖原豐，「因果關聯規則挖掘」，碩士論文，2006

[7]. 洪煌龍，「資料探勘在社區大樓電信服務市場之應用」，碩士論文，2004

[6]. 賴志鏹，「虛設行號案件調查蒐證之研究」，碩士論文，2005

[13]. 劉興浚，類神經網路在稅務稽徵選案模式之應用，元智大學管理研究所碩士論文，1998。

[23]. Kantardzic, M., “Data Mining: Concepts, Models, Methods, and Algorithms,” John Wiley & Sons, 2003.