From Global To Local: A Context-Embedded Lstm Recurrent Network For Local Content Popularity Prediction

With the emerging technology of mobile edge computing and the ever-increasing demand for low-latency multimedia services, it is intriguing research to predict online multimedia content popularity at the fine level of geographic granularity for potential mobile and edge services. However, such local popularity prediction poses new challenges, including data sparsity, the correlation between user behaviors and contextual information (e.g., geographical PoI distribution), making traditional regression-based popularity prediction inefficient. In this paper, we present a context-embedded local popularity prediction framework. First, based on measurement studies of 2 million users watching 0.3 million videos using an mobile video app, we reveal the characteristics of "local popularity": contextual information can significantly improve local popularity prediction; Second, we propose a context-embedded LSTM recurrent network, leveraging the correlation between the geographical context and content popularity to estimate a fine-grained content popularity; Finally, we carry our trace-driven experiments to show that our design can significantly improve the popularity prediction accuracy, by up to 35% against conventional regression models.