Random Subspace Methods with Automatic Selecting Parameter based on Kernel Smoothing

• Main Authors: Yu-Chen Hsieh, 謝友振
• Other Authors: Bor-Chen Kuo
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/10699344897558300034

碩士 === 國立臺中教育大學 === 教育測驗統計研究所 === 94 === In high dimensional data, how to solve small size problem is very important in many research domain. Random Subspace Method is a multiple classifier systems and has been shown that is a good approach to overcome small sample problems. In original Random Subspace Method, classifiers are constructed in the subspaces selected randomly from the original data space and the dimensionality of all subspaces is a fixed number. Then the decisions of base classifiers are usually combined by simple majority voting for the final decision. There is still not an effective way to estimate the best dimensionality of subspace, although it has a great impact on the classification result. In this paper, a weighted random subspace method with automatic subspace dimensionality selection has been proposed for classifying high dimensional data. The dimensionality selection method is based on the importance distribution of dimensionality estimated by kernel smoothing technique during the training process. Two feature weighting methods based on normalized re-substitution accuracy and Fisher’s linear discriminate analysis separability are introduced for improving the original subspace method. Experimental result shows that the proposed algorithm outperforms the original random subspace method.