Semi-supervised Classification For Brain-computer Interface

The thinking of semi-supervised learning is using a large number of unlabeled datascontaining the valuable informations to improve the performance with a small amount oflabeled datas. The semi-supervised classification as a research focus has had a broadprospects and is being applied to the biological medical field, for example, theclassification of the electroencephalogram, which is a key research of this paper. Theexisting semi-supervised classification approaches have two problems, which are thepoor training performance and misclassification during the labeling process. It isproposed that the corresponding solutions in this paper.Firstly, to solve the problem of poor training performance, the idea of lable meanwas applied to the designing of the semi-supervised algorithm. Two methods wereproposed which were Means4vm_mkl and Means 4vm_iter. We applied these twomethods to the three groups datasets of BCI Competition Dataset. Experiments showedthat both of the proposed algorithms achieved highly performances, especially on BCIⅠdataset. In addition, both of the proposed algorithms had high running speed laying afoundation for the online BCI system.Secondly, it was implemented in this paper that a novel marking strategy based onhelp training to avoid misclassification called semi-supervise linear surport vectormachine using help training, which was abbreviated as HTLSVM. There methods ofprobability density were applied in the stage of help training, which were called GMM，KNneighbor and Parzen window. We applied this method to the four groups datasets ofbenchark, which were g50c, BCIⅠ, BCIⅡ_Ⅱ, USPS, The results of the new methodwere better than S3VM and HTSVM on both of classification rate and efficiency. It wasproved that the new method proposed was effective.Thirdly, it was implemented that semi-supervise sparse classifier with combining thesparse classifier and fisher linear classifier based on the help training, and we used thesame data sets. In addition there was a detailed analysis on the g50c data set about simulation results.