Extending Attribute Information to Improve Classification Performance for Small Data Sets

• Main Authors: Chiao-WenLiu, 劉巧雯
• Other Authors: Der-Chiang Li
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/78366346167845175793

博士 === 國立成功大學 === 工業與資訊管理學系碩博士班 === 98 === Learning from small data sets is fundamentally difficult. In many data sets such as gene in medicine field or scheduling in the early manufacturing process, the data sizes are often not only small, but they also have high dimensions. Generally, a too small data size will detract modeling accuracy, and too many data attributes will affect the efficiency of the analysis. This research proposed a method for attribute analysis to enhance the analysis efficiency and accuracy for small data set. The proposed method includes two techniques; one called the class possibility method which uses a fuzzy membership function to build up the class possibility value for each data point in every attribute. The other technique called attribute construction aims to non-linearly create hidden attributes and combine attributes with high correlation value into principal attributes. Three data sets, an early flexible manufacturing system, Pima Indians diabetes data set, and Wisconsin breast cancer data set, are employed to prove the proposed method having better classification performance than other studies.