研究半监督学习方法在EEG分类中的应用.结合标签均值和自训练思想提出两种新的半监督支持向量机方法.首先通过未标记样本的预测值估计标签均值,然后对未标记样本的标签进行优化.在此基础上提出了两种半监督支持向量机方法,一种是基于多核学习的标签均值自训练半监督支持向量机(Means4vm_mkl)；一种是基于迭代优化的标签均值自训练半监督支持向量机(Means4vm_iter).对BCI Competition Dataset中的3组数据进行仿真实验,讨论分类正确率和运算效率两个指标.结果表明,两种方法均有较高的分类正确率,尤其在BCI Ⅰ数据集中,Means4vm_mkl方法达到了竞赛第一名的水平96％；而且运算效率较高,最快的只需29.5s,为在线BCI系统的设计奠定了基础.%In this paper the application of semi-supervised learning algorithms to brain-computer interface (BCI) was investigated. Two versions of semi-supervised support vector machines were proposed with integration of label mean and self-training. Firstly, the predictive values of unlabeled data were used to estimate label means, then the labels of unlabeled data were optimized by maximizing the margin between the label means. Based on these, two versions of the mean S3VM were proposed. One version is based on multiple kernel learning ( Means4vm_mkl) , the other one is based on alternating optimization ( Means4vm_iter). We applied these two methods to the three groups datasets of BCI Competition Dataset. Experiments showed that both of the proposed algorithms achieved highly performances, especially on BCI I dataset. The classification rate of Means4vm_mkl on BCI I dataset was 96% , the same as the topping one during the Competition. In addition, both of the proposed algorithms had high running speed with the shortest performance time of 29. 5 s, laying a foundation for the online BCI system.
展开▼
