网络流量识别的自适应分级滑动窗决策树算法

针对网络流量存在概念漂移、不同应用类型数据流偏态分布等特性, 提出了基于Hoeffding决策树的自适应分级滑动窗决策树的网络流量识别算法。该算法根据节点信息增益率检测概念漂移、动态调整概念漂移检测窗口及不同类型训练样本集窗口, 实现对不同速率概念漂移的自适应分类和决策树更新。实验结果显示新算法对劣势频繁漂移的应用类型的识别准确率与batch C4. 5算法接近, 比CVFDT算法提高约20%, 可以获得更加均衡的不同应用类型分类准确度。

Traffic identification algorithm based on Hoeffding decision tree with adaptive grading slide windows

Network traffic has characteristics of concept drift, unbalance distribution among different application types. This paper proposed a traffic identification algorithm, named adaptive grading slide window decision tree(AGSW-DT), based on Hoeffding decision tree. It realized adaptive detection of concept drift and decision tree update according to the information gain ratio of nodes, and then adjusted concept-drifting detection window and training set windows dynamically in accordance with the detection results. Comparing to the experiment results of batch C4. 5 and CVFDT, AGSW-DT algorithm gained approximate precision with batch C4. 5 algorithm and higher than that of CVFDT algorithm with 20% in terms of minor frequent concept-drifting application types. The proposed algorithm can obtain more balanced classification accuracy among different application types.