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.
展开▼
