基于分类面拼接的快速模块化支持向量机研究

On Pasting Small Fast Modular SVMs for Classification

针对大多数现有的机器学习算法处理大规模问题时需要的训练时间很长和存储空间很大的难点,提出了基于分类面拼接的快速模块化支持向量机算法(psfm-SVMs).在训练阶段,psfm-SVMs采用一簇平行超平面对大规模问题实施软划分,然后针对每个子问题并行训练支持向量机.在测试阶段,测试样本坐落于哪个子问题所在空间中,就由该子问题训练的支持向量机给出判别结果.在4个大规模问题上的实验表明:与采取硬划分的快速模块化支持向量机(fm-SVMs)相比,软划分能够使psfm-SVMs得到更加光滑的分类面,因而ps-fm-SVMs的泛化能力较高.在不增加训练时间的条件下,psfm-SVMs减少了由于训练集分割导致的分类器泛化能力下降.

Most of the current machine learning algorithms are very difficult in handling large scale classification problems for its long training time and large memory demand. This paper proposed an algorithm to paste small fast modular support vector machines （psfm-SVMs）. In the training phase, psfm-SVMs used a cluster of parallel super-planes to softly partition a large scale problem into many smaller sub-problems, and then many support vector machines were each trained in parallel with the sub-problems. In the test phase, at the location of the test sample, the classification was executed by the corresponding SVM. The experiment results on 4 large scale classification problems have illustrated that, compared with fast modular support vector machines （fm- SVMs） based on hard partitioning, soft partition could enable psfm-SVMs to get more smooth classification surface and higher generalization ability. Psfm-SVMs can cut down the descent of generalization ability but does not increase the training time.