Support Vector Domain Description with a new confidence coefficient
Mohamed El Boujnouni, Mohamed Jedra, Noureddine Zahid · 2014
Support Vector Domain Description (SVDD) has been introduced as a powerful technique for solving classification problems. It is a popular machine learning technique which tries to fit a hypersphere with minimal volume containing most of normal data, rejecting most of negative data. It can obtain more flexible data description by using suitable kernel functions. SVDD considers all data points with the same importance, consequently SVDD is very sensitive to uncertain data (noisy data or outliers), to deal with the uncertainty of data a confidence coefficient can be associated to each training sample. In this paper we propose a new method to generate those confidence coefficients. The experimental results show that our proposed approach significantly improves the classification accuracy.