基于Bootstrapping的因特网流量分类方法

发布时间：2018-05-03 07:18

本文选题：半监督学习 + 类别不平衡　；参考：《北京邮电大学学报》2014年05期

【摘要】：针对因特网流量分类面临的流量类别标记瓶颈和类别样本数分布不平衡,提出基于Bootstrapping的流量分类方法,使用少量有标记样本训练初始分类器,迭代利用无标记样本扩展样本集并更新分类器.在构建扩展样本集过程中,将无标记样本在某后验概率分布下的正确分类行为视为一个概率事件,建立新的置信度计算方法,以减少扩展样本集中的噪声样本;基于概率近似正确学习理论建立启发式规则,注重选择小类样本加入扩展样本集,缓解类别样本数分布的不平衡.实验结果表明,与初始分类器相比,基于Bootstrapping的流量分类器总体分类准确率可提高9.46%;与现有半监督学习方法相比,小类分类准确率提高2.22%.
[Abstract]:Aiming at the bottleneck of traffic classification and the unbalanced distribution of class samples in Internet traffic classification, a new traffic classification method based on Bootstrapping is proposed, in which a small number of labeled samples are used to train the initial classifier. Iteratively using unlabeled samples to extend the sample set and update the classifier. In the process of constructing an extended sample set, the correct classification behavior of unlabeled samples under a posteriori probability distribution is regarded as a probability event, and a new confidence calculation method is established to reduce the noise samples in the extended sample set. Heuristic rules are established based on probabilistic approximate correct learning theory, and small class samples are selected to join the extended sample set to alleviate the imbalance in the distribution of the number of samples. The experimental results show that compared with the initial classifier, the overall classification accuracy of the flow classifier based on Bootstrapping can be improved by 9.46%, and the accuracy of subclass classification can be improved by 2.22 points compared with the existing semi-supervised learning methods.
【作者单位】：华南理工大学软件学院;华南理工大学计算机科学与工程学院;
【基金】：国家自然科学基金项目(61171141)
【分类号】：TP393.06

【参考文献】