由於互聯網的便利性和電子商務技術的創新，信用卡支付數量的增長比以往還要更加強勢。信用卡交易量的急劇增加，同時詐欺交易的數量也同時增加。對金融業者來說，信用卡欺詐所造成的成本會使機構蒙受巨大損失，因此如何識別欺詐交易，甚至建立欺詐偵測系統 (FDS) 已經成為金融業者的主要問題之一。持卡人表現出特定的消費行為。每一位持卡人都可以用一組典型的購買類別模式表示，例如上次購買的時間，以及消費的金額等，但是該模式的偏差會對欺詐偵測系統產生潛在威脅。本論文探討如何將隱藏式馬可夫模型 (HMM) 運用到信用卡交易過程的序列中和偵測信用卡欺詐。應用 HMM 的優點在於它能夠即時偵測以及有效的增加準確率。本文提出一個可以同時檢測以單一持卡人為中心和以機構 (銀行或者是商家) 為中心的詐欺交易之框架。我們使用 BankSim 的資料進行單一持卡人與機構的詐欺偵測，透過 K-means 聚類演算法，將交易金額分為幾個觀察符號。先用訓練資料估計模型參數，再將所估計的參數帶入 HMM 測試資料。以機構為中心的交易，由於交易數量龐大，使用非重疊窗口偵測欺詐交易。單一持卡人為中心的交易，交易數量相對前者小許多，使用重疊的窗口來偵測欺詐交易。為衡量 FDS 的效能，文章中使用真陽率、偽陽率等指標進行效能判斷，並與其他分析方法比較，結果顯示採用 HMM 捕追詐欺的能力佳，但在準確率上略為失色。 Credit card payment has strongly growing due to easy access to internet and innovations in e-commerce technology. The number of transactions has increased dramatically not only on regular credit card consumption but also on fraud events. Identifying fraudulent transactions and establishing an efficient fraud detection system (FDS) have become major issues in the financial industry as costs of credit card can bring substantial losses for financial industry. By regarding the true-fraud transactions as hidden states, this thesis applies a hidden Markov model (HMM) to analyze the sequence of operation in credit card transaction processing and shows how it can be used for detecting fraud. An HMM is a doubly stochastic process with an underlying Markov process that is not directly observable but can be inferred by analyzing another set of stochastic process which produces the sequence of observations. In practice, testing FDS is difficult as banks usually do not agree to share their data with researchers as well as no benchmark data set are available. To evaluate the efficiency of the HMM-based FDS, we consider three criteria, namely, true positive rate, false positive rate and overall accuracy. Simulation results show that the proposed HMM-based FDS performs more efficiently in terms of true positive rate compared with false positive rate. Overall, the HMM-based FDS performs well in terms of overall accuracy.
