Abstract: In recent years, brain network has been widely used in the diagnosis and classification of brain diseases. Considering the uncertainty characteristics of brain, some researches applied uncertain graph to brain network modeling. In the research of uncertain brain networks, traditional feature extraction methods mostly used mean, variance, range and other methods to extract subgraph features, but they have the problems of poor generalization performance and the difference between subgraphs cannot be directly measured. Therefore, relative range is proposed as the new feature extraction method. The advantage of this method is that it not only takes into account the maximum differences between subgraph patterns, but also takes into account the differences between subgraph patterns, which can effectively avoid the disadvantages of traditional methods. The results show that compared with other feature extraction methods, the classification performance of the proposed method is significantly higher than that of the traditional method. Meanwhile, under different feature selection methods, the relative range shows good classification performance and strong generalization. This study provides an important reference for feature extraction of uncertain brain networks.

Key words: uncertain brain network, frequent subgraphs, feature selection, machine learning, classification

摘要： 近年来，脑网络被广泛应用在脑部疾病的诊断和分类中。考虑到大脑的不确定性特征，先前的研究将不确定图应用在脑网络建模中。在不确定脑网络研究中，传统的特征提取方法多采用均值、方差、极差等进行子图特征提取，但是它们存在泛化性能差以及子图间差异无法直接衡量的问题，进而影响分类准确率。因此，相对极差被提出作为新的特征提取方法。这一方法的优点在于既考虑到子图模式间的最大差异，又考虑到子图模式间的组间差异，可以有效避免传统方法的弊端。结果表明，相对极差与其他特征提取方法相比，其分类性能显著高于传统方法。同时，在不同的特征选择方法下相对极差表现出较好的分类性能，具有很强的泛化性。该研究为不确定脑网络特征提取方法提供了重要的参考意义。

关键词: 不确定脑网络, 频繁子图, 特征选择, 机器学习, 分类