The migration of popular Catch-up TV services to modern Over-The-Top (OTT) multimedia delivery infrastructures creates a wide set of scalability challenges which are commonly addressed using Content Delivery Networks (CDNs) relying on caching nodes close to users. The use of general-purpose caching nodes, tailored for generic web content, is far from optimal as it does not consider the particularities of Catch-up TV content, namely its dynamic popularity behavior, superstar effects, and relevance decay, as shown in existing scientific literature. Since caches are limited in size and are relatively small when compared to the whole catalog of available Catch-up TV content, which may contain tens of thousands of TV programs, it is crucial to make the most out of the available resources. To address these issues, this paper proposes a novel content-aware cache replacement algorithm, Most Popularly Used (MPU), capable of taking advantage of content demand forecasts built using machine learning models, to significantly outperform traditional cache replacement policies, such as Least Recently Used (LRU), Least Frequently Used (LFU), and First-In-First-Out (FIFO), and approach the optimal theoretical hit-ratio limits. MPU leverages millions of Catch-up TV request logs to validate its results under realistic conditions.
