在破產預測與信用評估領域對資料正規化與離散化的比較分析;Comparative Analysis of Data Normalization and Discretization for Bankruptcy Prediction and Credit Scoring

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/84065

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/84065

Title:	在破產預測與信用評估領域對資料正規化與離散化的比較分析;Comparative Analysis of Data Normalization and Discretization for Bankruptcy Prediction and Credit Scoring
Authors:	羅聖明;Lo, Sheng-Ming
Contributors:	資訊管理學系
Keywords:	正規化;離散化;破產預測;信用評估;機器學習;Normalization;Discretization;Bankruptcy Prediction;Credit Scoring;Machine Learning
Date:	2020-07-29
Issue Date:	2020-09-02 18:00:00 (UTC+8)
Publisher:	國立中央大學
Abstract:	在過往的破產預測以及信用評估領域中，有許多研究在前處理時進行正規化，然而大多研究僅採用單一種正規化方法進行實驗。為了瞭解正規化在破產預測與信用評估領域的適用性，本研究蒐集了四個信用資料集(Australia、Japan、Germany、Kaggle)與四個破產資料集(Bankruptcy、Japan、TEJ-Taiwan、USA)，搭配四種正規化方法，minMAX、MaxAbs、Standard、Robust，並以三種不同的分類器，K-Nearest Neighbor、Logistic Regression、Support Vector Machine進行分類，期望能了解不同正規化方法對於結果的影響。另外，有鑑於近年也有研究在正規化後進行離散化，因此本研究也進一步探討是否正規化搭配離散化能夠更提升準確率並改善效能，主要採用三種離散化方法，最小化描述長度原則(Minimum Description Length Principle，MDLP)、卡方分箱法(ChiMerge)、CAIM(Class-Attribute Interdependence Maximization)。本研究發現在整體平均下，正規化方法(MaxAbs、Standard、Robust)對於AUC及Type II具有正面影響。而正規化若進一步搭配CAIM或MDLP，對於AUC及Type II會有更進一步的提升。在所有實驗組合中，Robust搭配MDLP在三種分類器都會達到最佳的AUC，而Standard搭配MDLP則會有最佳的Type II結果。;In the field of bankruptcy prediction and credit evaluation, many studies implemented normalization in data pre-processing, but most studies only conducted with a single normalization method. So, in order to understand the applicability of normalization in the field of bankruptcy prediction and credit evaluation. We collected four credit datasets (Australia, Japan, Germany, Kaggle) and four bankrupt datasets (Bankruptcy, Japan, TEJ-Taiwan, USA), with four normalization methods (minMAX, MaxAbs, Standard, Robust) and using three kinds of prediction models (K-Nearest Neighbor, Logistic Regression, Support Vector Machine) to examine the prediction performance of normalization. In addition, some studies have performed discretization after normalization in recent years, so this study further explores whether discretization after normalization can improve accuracy and performance, we use three differents kinds of discretization methods, Minimum Description Length Principle (MDLP), ChiMerge and Class-Attribute Interdependence Maximization (CAIM). This study founds that under the overall average, the normalization methods (MaxAbs, Standard and Robust) has a positive effect on AUC and Type II. Moreover, if the normalization is further combine with CAIM or MDLP, the AUC and Type II will effectively improve. In all experimental combinations, Robust with MDLP will achieve the best AUC in three classifiers, and Standard with MDLP will have the best Type II result.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	194	View/Open

社群 sharing

Loading...