Network Traffic Classification (NTC) is a key piece for network monitoring, Quality-of-Service manage- ment and network security. Machine Learning algorithms have drawn the attention of many researchers during the last few years as a promising solution for network traffic classification. In Machine Learning, ensemble algorithms are classifiers formed by a set of base estimators that cooperate to build more com- plex models according to given training and classification strategies. Resulting models normally exhibit significant accuracy improvements compared to single estimators, but also extra time cost, which may ob- struct the application of these methods to online NTC. This paper studies and compares the performance of seven popular ensemble algorithms based on Decision Trees, focusing on model accuracy, byte accu- racy, and latency to determine whether ensemble learning can be properly applied to this modeling task. We show that some of the studied algorithms overcome single Decision Tree in terms of model accuracy and byte accuracy. However, the notable latency increase hinders the application of these methods in real time contexts. Additionally, we introduce a novel ensemble classifier that exploits the imbalanced pop- ulations presented in traffic networks datasets to achieve faster classifications. The experimental results show that our scheme retains the accuracy improvements of ensemble methods but with low latency punishment, enhancing the prospect of ensembles methods for online network traffic classification.
