DeepMEC: Mobile Edge Caching Using Deep Learning

• Published in: IEEE access Vol. 6; pp. 78260 - 78275
• Main Authors: Thar, Kyi, Tran, Nguyen H., Oo, Thant Zin, Hong, Choong Seon
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2018; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Base stations; Caching; Cloud computing; Computational modeling; Data centers; Data models; Deep learning; Mobile edge computing; Model accuracy; Neural networks; Popularity; prediction model searching; Prediction models; Predictions; Predictive models; proactive caching; Recurrent neural networks; Training; Videos
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2018.2884913

Caching popular contents at edge nodes such as base stations is a crucial solution for improving users' quality of services in next-generation networks. However, it is very challenging to correctly predict the future popularity of contents and decide which contents should be stored in the base station cache. Recently, with the advances in big data and high computing power, deep learning models have achieved high prediction accuracy. Hence, in this paper, deep learning is used to learn and predict the future popularity of contents to support cache decision. First, deep learning models are trained and utilized in the cloud data center to make an efficient cache decision. Then, the final cache decision is sent to each base station to store the popular contents proactively. The proposed caching scheme involves three distinct parts: 1) predicting the future class label of each content; 2) predicting the future popularity score of contents based on the predicted class label; and 3) caching the predicted contents with high popularity scores. The prediction models using the Keras and Tensorflow libraries are implemented in this paper. Finally, the performance of the caching schemes is tested with a Python-based simulator. In terms of a cache hit, simulation results show that the proposed scheme outperforms 38%, convolutional recurrent neural network-based scheme outperforms 33%, and convolutional neural network-based scheme outperforms 25% compared to the baseline scheme.