文献类型：期刊文献

英文题名：Mobile Service Traffic Classification Based on Joint Deep Learning With Attention Mechanism

作者：Li, Changbing[1];Dong, Chao[1];Niu, Kai[1];Zhang, Zhengzhen[2]

第一作者：Li, Changbing

通讯作者：Dong, C[1]

机构：[1]Beijing Univ Posts & Telecommun, Key Lab Universal Wireless Commun, Minist Educ, Beijing 100876, Peoples R China;[2]Beijing Union Univ, Smart City Coll, Beijing 100024, Peoples R China

第一机构：Beijing Univ Posts & Telecommun, Key Lab Universal Wireless Commun, Minist Educ, Beijing 100876, Peoples R China

通讯机构：[1]corresponding author), Beijing Univ Posts & Telecommun, Key Lab Universal Wireless Commun, Minist Educ, Beijing 100876, Peoples R China.

年份：2021

卷号：9

起止页码：74729-74738

外文期刊名：IEEE ACCESS

收录：;EI(收录号：20212310458103);Scopus(收录号：2-s2.0-85107149961);WOS:【SCI-EXPANDED(收录号:WOS:000673544400001)】；

基金：This work was supported in part by the National Key Research and Development Program of China under Grant 2018YFE0205501, and in part by the Key Program of National Natural Science Foundation of China under Grant 92067202.

语种：英文

外文关键词：Deep learning; Streaming media; Telecommunication traffic; Protocols; Computational complexity; Computational modeling; Aggregates; Mobile service traffic classification (MSTC); joint deep learning; attention mechanism

摘要：With the rapid development of mobile devices, smartphones have become the chief access to Internet and generated huge mobile service traffic. Mobile service traffic classification (MSTC) has been an important task that contributes to providing personalized services for end-users. With the excellent ability of automatic feature learning, deep learning has better performance than traditional machine learning methods. Giving more attention to a local focus, the attention mechanism can reduce computational complexity by filtering out useless information. Therefore, deep learning with attention mechanism can effectively realize automatic feature learning and reduce computational complexity. In this paper, a novel method for MSTC with a two-step strategy is proposed, which reduces the computational complexity of the deep learning model by attention mechanism. In the first step, a joint deep learning model is designed as a basic classifier, which learns features of mobile service traffic from multiple time scales. In the second step, the attention mechanism is adopted to aggregates the basic predictions generated in the first step. To verify this methodology, an experiment is performed to classify seven mobile services. The results show that we get the mean F1-score of 92.7% with 3.1 seconds time-delay, where the pure deep learning model gets the highest mean F1-score of 90.4% with 6.7 seconds time-delay.