With the development of Internet technology, it also brings an increasingly complex network security issues. The traditional traffic recognition technology such as port detection and deep packet inspection has been difficult to deal with the current increasingly complex network environment. As the theory of machine learning become mature, it has been successfully applied in many subject areas such as Image or voice recognition and medical fields. Machine learning methods simulate the human cognitive pattern by computers. The target of machine learning is to establish learning model by studying the existing knowledge and use the learning model to class or predict unknown data. In this research, machine learning methods were applied in internet traffic identification. Firstly, we introduced the research status of internet traffic identification and the relevant concepts of machine learning. Secondly, the main work is to research and compare the influence of different feature selection to identification accuracy based on three machine learning classification algorithms. The author proposed an improved feature selection algorithm and verified the effectiveness of this algorithm by experiments.