Intelligent Edge Learning for Personalized Crowdsourced Livecast: Challenges, Opportunities, and Solutions

Recent years have witnessed the expeditious development of crowdsourced livecast, also referred to as crowdcast, breeding such industry upstarts as Youtube Live, Twitch Tv, Mixer, Douyu, and so on. Unlike traditional TV-based livecast that provides uniform services, in crowd-cast, service providers are racking their brains to satisfy each viewer's personalized QoE demands since different viewers usually have quite diverse preferences. To achieve this goal, the key challenges lie in the accurate prediction of viewers' personalized QoE preferences and the cost-ef-fective viewer serving at the network edge. We argue that the confluence of edge computing and recent advance in deep learning shed light on accommodating these challenges, enabling more intelligent personalized QoE provision. In this article, we first provide an overview of today's crowdcast solutions, highlighting the challenges and opportunities therein from uniform crowdcast to personalized crowdcast with a large-scale measurement. We then present an intelligent edge learning framework ELCast as a case study, which leverages convolutional neural network and deep reinforcement learning in an edge computing architecture for personalized crowdcast. Trace-driven experiments demonstrate the superiority of ELCast over state-of-the-art approaches.