Abstract
Mobile traffic has grown very fast in recent years, particularly for videos watched on mobile devices. Cluster-centric small cell networks are used to serve the huge demand of watching videos. To avoid bottleneck in the limited capacity of backhaul link to the core network, caching in the network edge is such a way that the cached video could be delivered with less network traffic. The motivation of the prediction-based caching plan is to predict the most popular videos in a period, and cache them in advance for further requests. In the training phase, users who have similar preference to popular videos will be assigned to the same cluster, and the cluster will be assigned to a small cell for serving the clustered users. Some small cells could cooperate together as a cluster of small cells, and share their cache space with the help of distributed LT codes. During the serving phase, the most popular videos will be predicted, and they could be cached to serve requests in the future. New users will be assigned to small cells based on their preference. Clustering users with similarity could not only decrease the possibility of caches being replaced, but also has the benefit of predicting videos ranking more accurately. As the simulation result of one test case, the proposed methods decreased the global download rate by 34.2% with top 200 videos, 29.9% with top 50 videos, 28.6% with top 30 videos, 23.6% with top 10 videos compared to non-cooperative caching plan, and decreased by 7.4% with top 200 videos, 7.1% with top 50 videos, 5.8% with top 30 videos, and 2.6% with top 10 videos compared to cooperative but not predictive caching plan.