基于Bootstrapping的因特网流量分类方法

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%.

Study of Internet Traffic Classification Method Based on Bootstrapping

Aiming at the class labeling starvation and class imbalance problems in Internet traffic classification, a bootstrapping based traffic classification method was presented. An initial classifier was trained on a small number of labeled samples, and then it is updated iteratively by predicting the class labels of unlabeled samples and extending the training set. A new algorithm was devised to compute the confidence used for selecting new labeled samples into the extension set. It correctly adopts classifying unlabeled samples with a posterior probability distribution as probabilistic event and to decrease the noise in the extension set. Moreover, the heuristic rule was built with aid of probably approximately correct theory, its biases is toward selecting minority class samples into the extension set so as to reduce class imbalance degree. Experiments show that the bootstrapping based classifier gets improved of 9.46% on overall classification accuracy compared with initial classifier, and the recalls of minority classes get increased about 2.22% averagely compared with the existing method.