Data Reduction in Bankruptcy Prediction

• Main Authors: Kai-Chun Cheng, 鄭凱駿
• Other Authors: Chih-fong Tsai
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/45639385298165269954

碩士 === 國立中正大學 === 會計與資訊科技研究所 === 97 === Prior researches of using data mining techniques in bankruptcy prediction focus mainly on constructing effective prediction models. Particularly, many of them develop hybrid models based on feature selection in the pre-processing stage. However, very few of them emphasize on data reduction. Data reduction can make the training dataset cleaner and reduce outlier data, which can improve prediction accuracy. Therefore, the purpose of this thesis is to build up a data reduction method by using K-means to find the center of each cluster, and calculate the distance from all the data in a specific cluster to its cluster center. Then, we reduce the farther data in different percentages as the outlier data. We use four commonly used datasets in the bankruptcy prediction domain and employ neural networks, decision trees, logistic regression, and support vector machines as the prediction models after data reduction. The experimental results show that when the model trained by the four classifiers using four datasets after data reduction, the accuracy in general is higher than the model without data reduction. Moreover, the accuracy becomes higher when the reduction percentage increases.