Video Popularity Forecasting to Improve Cache Miss Rate in Content Delivery Networks

Abstract

Video transmission is of critical interest in several practical applications. Recent studies show that video content is highly cacheable in content delivery networks. Proactive and hybrid reactive-proactive caching policies, with the use of media popularity forecasting, are being developed as a better approach to conventional reactive cache strategies. Neural networks have been extensively used in popularity forecasting, however, training a neural network is a challenging NP-hard optimization problem. In this paper, we propose to train neural networks for video popularity forecasting with a novel continuation approach and Particle Swarm Optimization algorithm to improve forecasting accuracy. We create a dataset from an online video transmission platform and develop a cache simulation to find the relationship between forecasting accuracy and cache efficiency. Our findings support that higher accuracy have a significant effect in cache efficiency. Further results show that our neural network training approach is able to improve forecasting accuracy respect gradient based algorithms and therefore improve cache efficiency.