Impact of prediction uncertainty of popularity distribution on proactive caching

Proactive caching at wireless edge can reduce back haul traffic load or even offload wireless traffic with periodical popularity distribution prediction. The gain from proactive caching highly depends on the uncertainty of the prediction, which however is not well understood. In this paper, we analyze the impact of prediction uncertainty of dynamic popularity on probabilistic caching. To this end, we employ two neural networks for prediction, and consider the often-used MovieLens 1M dataset and a real dataset that can capture the request behavior of users for Youku videos for evaluation. To understand the impact of dynamic popularity on the prediction, the two real datasets are shuffled to obtain the corresponding static datasets. Simulation results show that the MovieLens dataset is near-static, which cannot capture the cold-start problem. The prediction uncertainty includes not only additive error, but also miss alarm and false alarm. The performance loss of proactive caching induced by the three types of prediction uncertainty differs for the two real datasets and differs for base stations with different cache sizes. For the MovieLens dataset, the additive error is the key factor that leads to the performance loss. For the Youku dataset, miss alarm incurs the most performance loss and additive error incurs the least loss when the cache size is large. For both datasets, the additive errors follow a summation of two Gaussian distributions.