中文摘要

線上拍賣蘊含龐大商機，但也吸引許多詐騙者混雜其中，有鑒於此，專家學者們紛紛提出提出各種詐騙偵測方法，以免無辜的消費者受害。在實務上，詐騙偵測的準確性與所使用的分類屬性集之特性息息相關，優良的偵測屬性集能以較少屬性，產生較佳的偵測效果。然而，前人大多使用經驗法則來設計屬性集，為免疏漏，應有更系統化、更周全的考量。有鑒於此，本計畫致力於發展詐騙偵測屬性集的挑選與建構方法，以提升詐騙偵測的準確性。首先，我們提出了一套基因式的屬性挑選方法，在演化過程中，除了準確率外，也顧及偵測成本的多寡，以產生一組低成本、高效能的詐騙偵測屬性集。此外，我們也發展一套語法演化式的屬性建構方法，以BNF為基礎，配合基因演算法，以各種不同方式組合原生屬性，以產生高效能的複合屬性。為了驗證提出方法的有效性，我們使用拍賣網站真實交易資料來進行實驗，結果顯示，本研究提出的方法能有效縮減屬性集的大小，並獲得較佳的準確率。 

English Abstract

In the recent years, the fast development of online auctions is obvious to all. However, it also attracts a lot of fraudsters to defraud in such a convenient platform. In view of this, researchers have proposed a lot of fraud detection methods to help the users identify the fraudsters. In practice, the effectiveness of fraud detection is strongly related to the chosen feature set. A proper feature set would result in building an effective detection models. To this end, this project developed effective feature selection and construction methods for online auction fraud detection. For feature selection, a genetic-based algorithm incorporated with a specialized cost function to obtain a compact, low cost but effective feature set. For feature construction, this project designed a BNF grammar evolution procedure to automatically generate more effective features. To evaluate these proposed methods, real transaction data gathered from auction sites is used for experiments. The result shows that the constructed feature set is actually helpful in building an effective detection model with low cost. 

資料來源：https://goo.gl/pib2vK