應用於分析信用評估巨量資料的決策樹分類法

Abstract

信用評估已成為金融機構評估是否核准顧客申請貸款的一項重要步驟。企業能根據不同的資料探勘技術來建立穩定而可靠的評估標準，降低放貸所產生的違約風險。但實際上，資料的來源已進入巨量資料時代。大量而雜亂的資料，加上複雜的資料探勘技術，大大提高了企業在資料處理與模型應用的難度。因此，本研究提出了一個使用決策樹作為主要的資料探勘模型來解決信用評估問題的方法（DTCAA）。其可讀性與良好的預測能力，以及它所產生的各種風險規則，有助於企業更好地理解顧客的特性，並能準確地結合實務應用。另外，本研究亦提出多種資料處理的方法來解決巨量資料下雜亂的資料定義與來源所帶來的問題，降低實務應用的門檻。經由使用真實的車貸申請資料，本研究驗證了DTCAA在實務上的可行性。即使在不同的違約比例與多種因素的改變下，決策樹同樣能夠提供與其他資料探勘方法相近的預測能力。

Credit assessment has been a large-scale problem among finance institutes. Their demand in reducing risk of customer debt can be achieved by applying data mining techniques to determine whether a new application should be approved or not. The problem, however, is actually under a Big Data environment. Complicated preprocessing steps are required because of the vast and messy data sources. The study proposes a Decision-Tree-Based Credit Assessment Approach (DTCAA) to solve the problem. Decision tree model is selected because of its interpretability and easily understanding rules, as well as its competitive performance. Additionally, the approach also includes various methods for data preprocessing. The consolidations can reduce messiness of the raw data, facilitating the implementation process. By acquiring the real data from one of the three biggest car collateral loan companies in Taiwan, the experiments indicate that decision Tree is competitive among various situations. Within multiple factors, the experiments suggest the usability of DTCAA in practice.