Research on new reduction strategy of SVM used to large-scale training sample set

摘要

It has become a bottleneck to use Support Vector Machine (SVM) due to such problems as slow learning speed, large buffer memory requirement, low generalization performance and so on, which are caused by large-scale training sample set. Concerning these problems, this paper proposed a new reduction strategy for large-scale training sample set. First authors train an initial classifier with a small training set, which is randomly selected from the original samples, then cut the vector which is not Support Vector with the initial classifier to obtain a small reduction set. Training with this reduction set, final classifier is obtained. Experiments show that the learning strategy not only reduces the cost greatly but also obtains a classifier that has almost the same accuracy as the classifier obtained by training large set directly. In addition, speed of classification is greatly improved.