Impacts of Learning User Request Behavior with a Real Dataset on Proactive Wireless-edge Caching

Since the wireless edge node has a small cache size, the performance of proactive caching is much better than passive caching when file popularity is known. Recent studies in the scenario where file popularity is unknown and need to be predicted found that proactive caching is still better than passive caching. However, most of the work is based on synthesized data sets or open source datasets collected in areas such as recommendation systems, which can barely reflect the request behavior of wireless users. In this paper, a dataset recording the number of requests for videos measured in a local area is used, based on which a neural network is employed to predict the short-term behaviors of individual and group user. The predicted user behavior information is then applied for proactive caching at macro or micro base stations. The research results show that when the measured dataset is adopted, the false alarm, missing alarm and additive errors, which are caused by strong temporal locality or even burstiness of user request behavior, make passive caching outperform proactive caching, especially for caching at macro base station. Once a synthesized static dataset is used, proactive caching performs significantly better than passive caching. This means that it is not sufficient to use only additive errors to characterize the uncertainty of popularity prediction, while reducing false and missing alarms is more important in order to achieve the performance gain of proactive edge caching.