Machine Learning Assisted Content Delivery at Edge of Mobile Social Networks

The demand of content delivery (for video streaming, photos, mobile web pages) has explosively grown in recent years, leading to an excessive volume of network traffic at the edge of mobile networks. Mobile social networking (MSN) is widely recognized as the most popular mobile application, as it attracts the largest number of users, and consumes the highest daily time of each user on average. However, mobile users can be distantly located, and their preferences to content may vary over time, leading to highly dynamic distribution of network traffic. Hence, it is necessary to allocate network edge resources for meeting the growing user demand to different types of contents spatially and temporally. To study this problem, we establish a machine learning framework for predicting how users' location and their consumed content can affect the traffic distribution in the underlying content distribution network. Given the predicted distribution of content demands, the proposed smart content delivery scheme is able to instruct the edge servers to dynamically deliver the content to edge users. We evaluate the performance of the conventional regression models and the proposed content delivery scheme over a real-world content distribution network for MSNs. Results reveal that the neural network model outper-forms other approaches by increasing the hit ratio by 30%, and saving the network cost by 15%.