基于主动学习的有监督在线多核学习算法

A Supervised Online Multiple Kernel Learning Algorithm Based on Active Learning

为了提高大规模数据的分类性能,提出一种基于主动学习的有监督在线多核学习算法SOMK_AL(Supervised online multiple kernel learning algorithm based active learning).首先,采用主动学习的方法缩减数据规模.通过训练生成两个分类器,对读入数据xt进行预测,将两个分类器预测类别不一致的数据作为信息含量高的有标记数据,参与在线学习过程中的核更新;接着,在核集成过程中,通过随机抽样的方法构造核函数集的子集,仅仅在子集中实现核更新,缩减核更新的计算规模.最后,在大规模数据的基准数据集上进行实验,对提出的算法的有效性进行评估,结果表明SOMK_AL能较好地提高数据的分类性能.

In order to improve the classification performance of large scale data, a kind of supervised online multiple kernel learning algorithm based on active learning （SOMK_AL） is proposed. Firstly, the data size is reduced by the active learning, in which the input sample xt is predicted by two classifiers and only the one predicted different is seen as the highest information sample to participate in the online learning process. Then, in the process of nuclear integration, the subsets of the kernel function set are constructed by random sampling and only the sample in the subset is to participate in the kernel updating to reduce the workload of kernel update. Finally, experiments are carried out on the large-scale benchmark data sets to evaluate the effectiveness of the proposed algorithm. The results show that the proposed SOMK_AL could improve the classification performance.