隨著金流與物流等基礎建設的成熟，電子商務的蓬勃發展有目共睹，不但已成為現代人生活一部分，交易金額也年年攀高。在2017年，全球電子商務總銷售額已高達2.29兆美元，其興盛程度可見一斑。但面對如此龐大的交易金額，也引起不肖人士的覬覦，在電子商務平台中進行詐騙，而線上拍賣詐騙更佔其中的大宗。有關線上拍賣詐騙偵測，已有許多方法被提出，但對於日新月異的詐騙手法，其準確率仍有待提升。為解決此問題，本研究將配合模型融合概念，發展有效的詐騙偵測方法。首先，我們以線性迴歸組合數種傳統的分類模型，以產生更有效的融合模型，並比較傳統單一分類模型與融合模型之間的差異。之後，以不同訓練資料配比，將產生各種不同特性之模型，以多階連續過濾以及平衡過濾方式加以整合，以提升詐騙偵測的準確性。此外，由於偵測屬性集與偵測效能息息相關，本研究也探討屬性篩選對於偵測準確率之影響。為驗證提出方法之有效性，本研究採用Yahoo!奇摩實際交易資料進行實驗。與四種單一偵測模型相比較，結果顯示融合模型確實能提高偵測準確率。當使用連續過濾與平衡過濾流程時，除能獲得高準確率外，也能分段獲得較高之偵測精度。此外，結果亦顯示，使用Principle Component Analysis或Wrapper法進行屬性篩選，並無助於結果的改善。由上述結果可知，本研究提出方法確有助於改善詐騙偵測準確率，提供消費者更周全的購物安全防護機制。

With the maturity of infrastructure such as cash flow and logistics, the booming development of e-commerce is obvious to all. Not only has it become a part of modern life, but the transaction amount has also increased year by year. In 2017, global e-commerce total sales have reached 2.29 trillion US dollars, and its prosperity can be seen. However, in the face of such a large transaction amount, it also attracts a lot of fraudsters to join the e-commerce platform. Among the reported cases, online auction fraud undoubtedly forms a large proportion. There have been many methods for online auction fraud detection, but the accuracy of the ever-changing fraud scheme still needs to be improved. In order to solve this problem, this study adopts the model fusion concept to develop effective fraud detection methods to improve the accuracy of detection. First, we combine several traditional classification models with linear regression to produce a more efficient fusion model and compare the differences between the traditional single classification model and the fusion model. After that, training sets with different fraud/non-fraud ratio are used to build detection models of different characteristics. Based on these models, a multi-level continuous filtering and a balanced filtering method are developed to integrate these models and improve the accuracy of fraud detection. In addition, since the detection attribute set is closely related to the detection performance, this study also explores the impact of attribute screening on detection accuracy. In order to verify the validity of the proposed method, the study used Yahoo!Kimo actual transaction data for experiments. Compared with the four single detection models, the results show that the fusion model can improve the detection accuracy. When using continuous filtering and balanced filtering processes, in addition to high accuracy, segmentation can achieve higher detection accuracy. In addition, the results also show that feature selection does not contribute to the improvement of the results. From the above results, the proposed method does help to improve the accuracy of fraud detection and provide consumers with a more comprehensive shopping security protection mechanism.

第一章 緒論	1
第二章 知識背景與技術介紹	5
2.1 線上拍賣詐騙	5
2.2 模型融合(Model Fusion)	5
2.3 分類方法	9
第三章 以多模型融合為基礎之詐騙偵測方法	14
3.1詐騙偵測屬性集	14
3.2 以模型融合建立詐騙偵測模型	16
3.3 多階連續過濾之詐騙偵測流程	19
3.4 偵測流程	25
第四章 實驗結果與討論	28
4.1 實驗設定	28
4.2 線性融合模型之效能測試	29
第五章 結論與未來工作	34
參考文獻	36
附錄A: 屬性篩選結果	38

表2.1常見詐騙類型	5
表3.1本研究使用之37種詐騙偵測屬性	14
表3.2各種分類器對於詐騙偵測之效能比較	16
表3.3使用PCA與Wrapper法進行屬性挑選之偵測結果*	17
表3.4以線性方式融合4種分類器之偵測準確率	19
表3.5在不同NF:F配比下塑模之偵測結果(Random Forest)	20
表3.6在不同NF:F配比下塑模之偵測結果(AdaBoost)	21
表4.1 Confusion Matrix	28
表4.2單一模型與線性融合模型之偵測準確率比較	30
表4.3過濾式以及平衡式融合模型與RF之準確率比較	31
表4.4多階連續過濾以及平衡過濾融合模型與RF之準確率比較(PCA篩選)	32
表4.5連續過濾偵測流程之各階段偵測精度	33

圖2.1(a) 多層偵測概念	6
圖2.1(b) 應用多階段偵測之研究	7
圖2.2(a) 基本互補式融合流程	8
圖2.2(b) 應用互補融合流程之研究	8
圖2.3: 使用融合演算法進行模型融合	9
圖2-4(a) 決策樹	11
圖2-4(b) 隨機森林	12
圖3.1: 各種分類器在各種效能指標之比較	17
圖3.2: 詐騙偵測模型融合方法	18
圖3.3: 使用Random Forest配合不同NF:F配比之偵測效能變化圖	20
圖3.4: 使用AdaBoost配合不同NF:F配比之偵測效能變化圖	22
圖3.5: 運用多階連續過濾進行詐騙偵測	23
圖3.6: 平衡式過濾偵測流程	24
圖3.7: 多階連續過濾詐騙偵測流程	26
圖3.8: 平衡過濾詐騙偵測流程	27
圖4.1: 單一模型與線性融合模型之偵測準確率比較圖	30
圖4.2: 線性融合模型偵測結果：屬性篩選與未篩選之比較	31
圖4.3: 多階連續過濾及平衡過濾偵測流程與Random Forest之準確率比較	32
圖4.4: 多階連續過濾及平衡過濾偵測流程與Random Forest之準確率比較(PCA)	32

[1]	Alford M. (2013), "Intelligent fraud detection: a comparison of neural and Bayesian methods," Computer Fraud & Security, pp. 14-16 
[2]	J.-S. Chang, W.-H. Chang/Electronic Commerce Research and Applications 13 (2014) 79–97
[3]	Chau, D.H., and Faloutsos,C. (2005). Fraud detection in electronic auction. European Web Mining Forum at ECML/PKDD
[4]	Chau, D.H., Pandit,S., and Faloutsos,C. (2006). Detecting fraudulent personalities in networks of online auctioneers. Proceedings of PKDD 2006, pp.103-144.
[5]	Chau, D.H., Pandit,S., Faloutsos,C., and Wang,S..:NetProbe:A fast and scalable system for fraud detection in online auction networks. Proceeding of the 16th International Conference on World Wide Web, pp. 201-210.(2007)
[6]	Chen C., et al. (2013), "Web Media Semantic Concept Retrieval via Tag Removal and Model Fusion," ACM Trans. on Intelligent Systems and Technology, Vol. 4, No. 4, Sep. 2013.
[7]	Chen, J., et al. (2015), "Big Data based fraud risk management at Alibaba," The Journal of Finance and Data Science 1 (2015) 1-10. 
[8]	eMarketer，Global Ecommerce Platforms 2017:A Country-by-Country Review of the Top Retail Ecommerce Sites(July 13,2017)，https://www.emarketer.com
[9]	Huang, S., et al. (2015), "A Hybrid Multigroup Coclustering Recommendation Framework Based on Information Fusion," ACM Trans. on Intelligent Systems and Technology, Vol. 6, No. 2, Mar. 2015.
[10]	Gavish, B., and Tucci, C. (2008), 'Reducing Internet Auction Fraud', Communications of the ACM, vo. 51, no. 5 , pp. 89-97.
[11]	Goes, UP., Tu, Y. and Tung, A.,:Online Auctions Hidden Metrics, Communications of the ACM, vol.52, No.4, pp.147-149.(2009)
[12]	Kim, K., Choi Y., and Park J. (2013), "Price fraud detection in online shopping malls using a finite mixture model," Electronic Commerce Research and Applications 12 (2013) 195-207.
[13]	National White Collar Crime Center(NW3C). 2017 Internet Crime Report. Retrieved on Oct. 1, 2017, from Internet Crime Complaint Center: http://www.ic3.gov/media/annualreport/2017_IC3Report.pdf
[14]	Mitchell,T. and McGraw-Hill, “Machine Learning”, 1997, pp.52-81
[15]	Pandit, S. et al., (2007). Netprobe: a fast and scalable system for fraud detection in online auction networks. Proceedings of the 16th international conference on World Wide Web (pp. 201-210). ACM.
[16]	Tsang, S., et al. (2014), "SPAN: Finding collaborative frauds in online auctions" Knowledge-based systems 71 (2014) 389-408.
[17]	Leo Breiman(2001), "Machine Learning",Volume 45,Issue 1,pp.5-32
[18]	Miao Yufei & Zhang Xiaohong(2015), "Improvement and application of C4.5 decision tree algorithm",2015,51(13):255-258
[19]	Guo Qiao-jin & Li Li-bim & Li Nig,"modified AdaBoost algorithm for imbalanced data classification",2008,44(21):217-221
[20]	詹佳憲、陳嘉平，「以多層感知器辨識情緒於國台客語資料庫」，The Association for Computational Linguistics and Chinese Language Processing, ROCLING 2016, pp. 10-21
[21]	劉祐宏，「線上拍賣詐騙偵測之屬性挑選與流程設計」，淡江大學資訊管理學系，碩士論文，民101
[22]	洪儀玶，「具早期預警能力之線上拍賣詐騙偵測」，淡江大學資訊管理學系，碩士論文，民96
[23]	鄭孝儒，「線上拍賣潛伏期詐騙者之有效偵測」，淡江大學資訊管理學系，碩士論文，民100
[24]	梁賀翔，「一套線上拍賣詐騙即時偵測系統」，淡江大學資訊管理學系，碩士論文，民99
[25]	林敬堯，「一套有效率的複合式線上拍賣詐騙偵測系統」，淡江大學資訊管理學系，碩士論文，民103
[26]	謝金育，「結合貝氏網路與激勵理論之推薦機制-電影推薦系統設計」，碩士論文，民102
[27]	內政部警政署，警政署統計室，警政統計通報(106年10月28日)，取自
https://www.npa.gov.tw