关键词: 并行支持向量机, 大规模数据集, 有限资源, 随机傅里叶特征, 一致中心调节

Abstract: Support Vector Machines（SVMs） have become popular classification tools, but when dealing with very large datasets, SVMs need large memory requirement and computation time. Therefore, large-scale SVMs are performed on computer clusters or supercomputers. A novel parallel algorithm for large-scale SVM is presented. The algorithm is performed on a resource-limited computing environment and guarantees a uniform convergence. The infinite-dimensional implicit feature mapping of the Gaussian kernel function is sufficiently approximated by a low-dimensional feature mapping. The kernel SVM is approximated with a linear SVM by explicitly mapping data to low-dimensional features using random the Fourier map. The parallelization of the algorithm is implemented with a consensus centre adjustment strategy. Concretely, the dataset is partitioned into several subsets, and separate SVMs are trained on processors parallel with the subsets. When the optimal hyperplanes on subsets are nearly found, solutions achieved by separate SVMs are replaced by the consensus centre and are retrained on the subsets until the consensus centre is optimal on all subsets. Comparative experiments on benchmark databases are performed. The results show that the proposed resource-limited parallel algorithm is effective and efficient.

Key words: parallel Support Vector Machines（SVM）, large-scale datasets, limited resource, random Fourier features, consensus centre adjustment