An ML-Based Approach for Near Real-Time Content Caching

Abstract

Content caching is a well-known promising solution to address large demands for streaming companies. This paper presents an ongoing work towards improving CDN network traffic focusing on users' quality of experience (QoE) by anticipating which videos will be popular on Globo's platform. To do so, a deep neural network approach was chosen to model video's popularity based on its metadata and a near real-time framework is presented describing how to make content caching in a preemptive way. Additionally, a threshold selection approach is presented defining whether a video should be cached or not. The presented approach allows making content cache without any user interaction, aiming to decide about the admission of the content before it starts to receive requests. This approach is important to most of the daily published videos at Globo, especially for breaking news. Using Globo's real-world data, we demonstrate the popularity predictor results and conclude with some directions for future works.