由於網路拍賣平台具有高匿名程度、不完善的法律規範、低度的市場進入障礙等因素,使得詐騙者可以輕易的在網拍活動中進行詐騙行為。在美國,網路拍賣詐欺造成的損失曾經達到所有網路詐騙損失金額的一半;在臺灣,拍賣詐財更在官方財務損失排行榜中名列前茅。故本研究以臺灣Yahoo! 奇摩拍賣為資料來源,利用社會網路分析 (Social Network Analysis, SNA) 的概念為基礎,提出創新的相似度衡量指標,根據帳號之間的交易關係評估帳號間之交易行為相似程度。並且以基因演算法 (Genetic Algorithm, GA) 結合常見的C4.5, SVM, NB等三種不同分類方法各自建立預測模型。根據5-fold交叉驗證之測試結果顯示,GA/NB與GA/SVM在不同資料集中,異常類帳號之召回率 (Recallabnormal) 最高可達100%,證明研究方法確實有效。
Because of the anonymity, loose constraints, and low entry costs associated with Internet auctions, fraudsters can easily set up Internet auction scams. Internet auction had accounted for almost half the Internet fraud in the US. In Taiwan, the auction fraud also caused high losses. This paper uses experimental data gathered from Yahoo! Auction Web site in Taiwan, and designs four similarity indicators to evaluate the transaction behavior similarity between whole accounts based on social network analysis concepts. Also, this study proposes the hybrid of genetic algorithm and Classification approach to construct abnormal accounts predict models that focus on the type of failure to ship. The testing results based on 5-fold average show the recall rate of abnormal reached 100%. Using GA/NB and GA/SVM can achieve superior classification performance.