RangeTree: A Feature Selection Algorithm for C4.5 Decision Tree

Abstract

In order to conduct fine-grained network management in mobile network, Traffic Classification or Detection is widely used to divide network traffic into different classes, according to their source applications. Many techniques are exploited in Traffic Classification. Among them, machine learning has grown considerably attention because of its accuracy. Feature selection chooses feature combinations for machine learning algorithms, and has significant influence on the accuracy and efficiency. To discover optimal features, all possible combinations need to be evaluated by testing real classifiers. With numerous features, feature selection can cost an abundance of time and computational resources. This paper proposes a feature selection algorithm for C4.5 Decision Tree. This algorithm utilizes structural characteristics of C4.5 algorithm to exclude some of the combinations without actually testing the classifiers. The simulation results demonstrate that the algorithm can reduce the number of tests in seeking the optimal feature combination.