Heuristic sample reduction method for support vector data description

摘要

Support vector data description (SVDD) has become one of the most promising methods for one-class classification for finding the boundary of the training set. However, SVDD has a time complexity of $O;(N^{3})$ and a space complexity of $O;(N^{2})$. When dealing with very large sizes of training sets, e.g., a training set of the aeroengine gas path parameters with the size of $N>10^{6}$ sampled from several months of flight data, SVDD fails. To solve this problem, a method called heuristic sample reduction (HSR) is proposed for obtaining a reduced training set that is manageable for SVDD. HSR maintains the classification accuracy of SVDD by building the reduced training set heuristically with the samples selected from the original. For demonstration, several artificial datasets and real-world datasets are used in the experiments. In addition, a practical example of the training set of the aeroengine gas path parameters is also used to compare the performance of SVDD based on the proposed HSR with conventional SVDD and other improved methods. The experimental results are very encouraging.