近年來,隨著社會風氣的改變,消費大眾的行為也日益不同,以現金交易的付款方式,逐漸地被信用卡所取代,可見信用卡的重要性日漸提高。然而,隨著這些改變,伴隨而來的卻是層出不窮的信用卡盜刷問題。在過去有關信用卡盜刷偵測的研究方法中,大都是利用他人大量的歷史消費資料來建立模型,去偵測某位使用者的盜刷情況。本研究則採用了一種有別於過去的個人化方法(Personalized Approach)來偵測盜刷。此法可在消費者擁有少數的真實交易資料時,或是在消費者尚未申請信用卡前,即可建立個人化模型來預防盜刷。個人化方法雖然提供了一種不錯的解決方案,但仍有許多問題亟待解決。舉例如下:(1)在蒐集消費資料時,消費者多半不願花太多時間去填答問項,導致蒐集之個人消費資料量過少。(2)由於動態的消費者行為或是消費者可能不願用心填答而產生資料矛盾的情形。 因此,本研究的主要重點在探討矛盾情況對預測準確率的影響程度,且在有限的交易資料下,觀察資料分佈的情況對預測準確率的影響,並設法提高盜刷的預測準確率。另外本研究也利用支持向量機、倒傳遞網路、以及二元支持向量系統來建立一個有效的信用卡盜刷偵測模型。研究成果顯示,支持向量機與倒傳遞類神經均可得到不錯的訓練結果,但在預測未來資料上,較高的自我訓練結果反而有較差的預測未來能力。除此之外,本研究也運用許多技巧來提高預測的結果,如過量取樣(Oversampling)、階層式(Hierarchical)SVM、投票多數法(Majority Voting)…等。結果顯示,上述幾種方法均可達成高的異常偵測率。另外,在工具的比較上,三種工具得到的結果差異不大,但BSVS是最簡易操作的工具。
Credit cards are a popular tool for transactions in many countries lately. However, credit card frauds have occurred frequently. How to detect credit card frauds, therefore, has become a key issue in recent years. Many previous studies proposed models which were constructed from the past real transaction data of many others to detect new transactions of a certain individual. In contrast to those traditional approaches, this study employs a personalized approach to solve the problem of credit card fraud. The personalized approach proposes to prevent fraud before the consumer uses a credit card or when the collected data are few. This new approach is promising. However, there are still some problems which have to be solved. For examples, 1) consumers are not willing to spend too much time to answer questions so that the collected data are few, 2) the dynamic consumer behavior may cause data overlapping. To improve the problems mentioned above, this research employs the personalized approach to address the credit card fraud problem. The main purpose of this study is to investigate the influences of data distribution on the prediction accuracy. Support vector machine (SVM), back propagation network (BPN), and binary support vector system (BSVS) are used to construct detection models for credit card fraud. The experimental results show that SVM and BPN can obtain good training results. However, both techniques fail to predict future data accurately for those cases with high training results. Besides, the classification results of these three classifiers are comparable. Compared to the other two techniques, BSVS is the easiest tool to use. This study also employs several techniques, such as hierarchical SVM, majority voting, and over-sampling, to improve true negative rates. Results from the experiments indicate that these techniques can increase true negative rates effectively.
為了持續優化網站功能與使用者體驗,本網站將Cookies分析技術用於網站營運、分析和個人化服務之目的。
若您繼續瀏覽本網站,即表示您同意本網站使用Cookies。