Abstract: Most of the present brain network classification methods are to construct a brain network by collecting the signals, and then it is classified according to one brain network feature attribute between one or more brain regions. So other attributes of the brain network in the classification are ignored, but these attributes likely have a greater impact on the experimental results. In order to solve these problems, this paper proposes a brain network classification method based on multi-form eigenvectors and new graph kernels. The classification method consists of four steps：firstly, the experimental original data is preprocessed to complete the brain network construction; secondly, according to different thresholds, various brain network attribute values are extracted in the brain network; thirdly, the support vector machine is used to train all data, according to the quality of the training results, the optimal threshold parameters are selected for each network attribute value, the feature combination is carried out. Finally, the eigenvectors trained by support vector machine. Through the analysis of experimental data and comparison with existing methods, the effectiveness of this method for brain network classification in mild cognitive impairment dataset is verified.

Key words: brain networks, support vector machine, Mild Cognitive Impairment（MCI）, graph kernels

摘要： 目前已有的脑网络分类方法大多是通过处理收集的信号来构建脑网络，并根据一个或多个脑区之间的脑网络特征属性来进行分类。该分类方法只考虑一个特征属性，忽略了脑网络的其他特征属性，而被忽略的特征属性很可能会对实验结果产生较大的影响。为了克服已有分类方法的缺陷，文中考虑多种特征属性提出了一种基于多形式特征向量的脑网络分类方法并使用了新型图核，该分类方法由4步构成：将原始实验数据经过预处理后完成脑网络构建；根据不同的阈值来提取脑网络中多种脑网络属性值；利用支持向量机训练所有数据，根据训练结果的优劣，在每种网络属性值里挑选分类效果最优的阈值参数，并将它们进行特征融合；使用支持向量机训练融合后的特征向量。通过实验数据分析并与已有分类方法进行了对比，验证该方法在轻度认知障碍数据集上脑网络分类的有效性。

关键词: 脑网络, 支持向量机, 轻度认知障碍, 图核