Multiple Classifier System for High Dimensional Data Classification

• Main Authors: Chia Hao Pai, 白家豪
• Other Authors: Bor-Chen Kuo
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/69456276393621271551

碩士 === 臺中師範學院 === 教育測驗統計研究所 === 92 === Traditional pattern recognition system depends on a single classifier. However, when data are highly dimensional and the training sample size is small, the classifiers may become weak and unstable. Multiple classifier system (MCS) is one of the solutions for improving weak classifiers. Bagging, boosting and the random subspace method are three popular techniques for multiple classifier system. They manipulate training data by some resampling and weighting techniques to produce complementary classifiers, and then combine them into a powerful decision rule. A great part of papers investigate the performance of MCS in original feature space. However, in high dimensional data classification, feature extraction is an important part for reducing dimensionality and mitigating Hughes Phenomenon. In this study, the performances of MCS with base classifier, quadratic classifier, nearest mean classifier and pseudo-fisher support vector classifier are investigated based on small training sample size problem in PCA, DAFE and NWFE feature space. The experimental result shows that new hybrid algorithms outperform bagging, boosting and the random subspace method.