Device to Device Caching Delivery Using Predicted Demand on Trajectory
Makoto Tsunekiyo, Noriaki Kamiyama · 2023
As video viewing on mobile terminals becomes more common, there is concern that the backhaul traffic load on cellular networks (CN) will increase dramatically. To reduce the backhaul load, mobile edge computing, which distributes video content from a cache at the base station, has been attracting attention. Moreove, another effective method to further reduce the load is to cache the content at the mobile terminal and distribute it via D2D（device-to-device）communication. However, since the cache capacity of the mobile terminal (MT) is limited, it is effective to preferentially cache content that is expected to be in high demand along the MT’s route of travel.In this paper, we propose content demand estimation using deep learning to D2D cache delivery. We propose a method to select contents to be cached on MTs by estimating contents that are likely to be demanded by other MTs on the travel route using a long-short term memory (LSTM) neural network, which is one of the algorithms of deep learning. First, we generated time-series data based on the number of keyword searches (number of viewings) for 10 well-known movie titles to confirm the effectiveness of the demand estimation part of the proposed method and the generality of the learning model. Next, a cache is created by making delivery requests based on the pre-movement demand distribution using the predicted values, and the hit rate with the content in the cache is calculated when delivery requests are made based on the post-movement demand distribution. Then, we compare the total number of requests per content measured and predicted for California (CA) and New York (NY) in the U.S, and we confirm that high-demand content can be predicted at the destination where the MT moves. Moreover, we evaluate the cache hit rate with and without considering the popularity bias of the content used in the simulation.