Using Data Mining Approaches to Construct a Integrated Credit Scoring Model

[[abstract]]信用卡產業面臨強烈競爭與呆帳問題，因此信用評等模型的重要性日益增加。近幾年，許多相關領域，例如統計(Statistics)、機器學習(Machine Learning)、人工智慧(Artificial Intelligence)，紛紛投入信用評等的研究。本研究目的為減少人為因素、節省作業成本及加強信用卡授信風險管理，故建立一套客觀、快速的信用卡授信方法與開發潛在客戶，以增加銀行收入並降低潛在之信用風險。 然而，過去的研究均只著重在建立一個高準確率信用卡分類或行為分類的模型，且各種分類器各在不同的情況下各有不同的表現，沒有任何一種分類器其分類結果永遠最佳，故本研究提出整合型信用評等模型，藉以探討何種模式為最佳。並藉由支援向量機模型(Support Vector Machines, SVM ) 與類神經網路(Artificial Neural Networks, ANNs)、案例式推理(Case Based Reasoning, CBR) 建立起信用評等模型。 ICSM分為兩步驟，第一階段為建構以支援向量機與類神經網路為基礎的信用評等模型，用以將申請人分類為信用優良與信用不佳兩類別。並比較 ANNs 與 SVM 的判斷結果是否一致，當二者判斷相同時，表示其診斷結果；若二者判斷發生不一致時，便進入第二階段。第二階段為建構以案例式推理為基礎的再分類信用評等模型，用以將被第一階段分類為不一致案例重新分配，以降低分類時可能產生之型一與型二誤差。[[abstract]]The use of credit scoring model has became a very important issue as the credit industry is competitive and has been experiencing many problems with big debts, therefore credit scoring model has been widely studied in the area of statistics, machine learning, and artificial intelligence during the past few years. The purpose of this study is to reduce the human factors, save operating costs, and reinforce credit risk management. We will construct an objective and fast credit scoring model and develop potential customers in order to increase income and reduce potential bank credit risk. However, Past studies only focused on establishment of a high accuracy of credit scoring model or behavior classification model, so that various classifiers have different performance in different database, therefore this study will present the integrated credit scoring model (ICSM) and explore the best mode. This paper constructs a credit scoring model by applying a hybrid modeling procedure with SVM, ANNs, and CBR. ICSM involves two stages. The first stage is constructing an SVM- and ANN-based credit scoring model, which classifies applicants with accepted (good) or rejected (bad) credits. Comparing the results of ANNs and SVM, if the judgment is the same what represent the diagnosis results, and if not, it will proceed to the second stage.The second one is trying to reduce the Type I and Type II error by reassigning the inconsistent applicants by using the CBR-based classification technique