SOMNN-SVM: A reduction strategy for large scale training sample set

Concerning a problem that convergence of NN-SVM is slow and outliers bring invalidation of this algorithm while training a large-scale sample set, SOMNN-SVM is presented in this paper. This algorithm removes outliers and constructs a small-scale representative sample set removing most of non-support vectors. Then in this small-scale set, vectors are retrimmed. Since NN algorithm is searching data in a small-scale sample set, the frequency of search is reduced and the speed of the convergence is boosted. Experiments have shown that SOMNN-SVM has ability to construct a subset which reflects most of the information from original large-scale sample set.