Feature Selection in the Classification of Network Traffic Flows

Abstract

The article provides analysis and comparison of sequential feature selection methods for training machine learning models intended to classify network traffic flows. Feature selection plays an important role in classification. First of all, feature selection helps to reduce the duration of model training in the future (by introducing fewer features in the data set); secondly, it improves accuracy of classification in some cases. As network traffic of contemporary multiservice networks is extremely varied, traffic flows feature a large number of statistical characteristics, which may be used as features in training data sets. Training of classifier models using all the features and the process of identification of application layer protocols in flows itself may take a rather long time. Feature selection helps to significantly reduce the training period for these models. The article describes sequential feature selection methods, such as Sequential backward selection, Sequential forward selection, Sequential forward floating selection, Sequential backward floating selection.