Decremental multi-output least square SVR learning

摘要

The solution of multi-output LS-SVR machines follows from solving a set of linear equations. Compared with ε-intensive SVR, it loses the advantage of a sparse decomposition. In order to limit the number of support vectors and reduce the computation cost, this paper presents a decremental recursive algorithm for multi-output LS-SVR machines. This algorithm removes one sample one time and large-scale matrix inverse is computed quickly based on previous results. The decremental algorithm can be used to train online multi-output LS-SVR machine. Experimental results demonstrate the effectiveness of the algorithm.