A New Boundary Distance Measure via Support Vector Domain Description and Its Application to Classification

• Main Authors: Li-Chun Chen, 陳麗君
• Other Authors: Shu-Mei Guo
• Format: Others
• Language: en_US
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/96281512889074475616

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 96 === A new boundary distance measure based on support vector domain description (SVDD) is proposed in this thesis, which measures the distance between an object and the boundary described by a few number of training objects, namely support vectors, in the SVDD. In the first part of this thesis, the boundary distance measure is derived by locating an object’s nearest boundary point with the help of geometrical interpretation of inner product, and then a simple post-processing method using the boundary distance measure is proposed, which tries to modify the SVDD boundary in order to achieve tight data description. The experimental results show that the proposed decision boundary can fit the shape of synthetic data distribution closely and can achieve better or comparable classification performance on real world datasets. Compared to kernel whitening which is often applied to improve the slack boundary issue with the SVDD, the proposed method can construct a better decision boundary more efficiently. In the second part of this thesis, the boundary distance measure is utilized for multiclass classification problems. To decide which class a test object belongs to, a minimum distance strategy for combining a multiple of SVDDs is then carried out. Experimental results also show that our proposed minimum distance algorithm can achieve a comparable or better performance than other support vector based classifiers.