Online Prediction of Video Popularity in OVSs: A Video Age-Sensitive Model With Beyond Views Features

Current Online Video Services (OVSs) usually optimize service strategies based on future video popularity to improve efficiency, making popularity prediction highly important. Existing models generally infer a video’s future views from its early views, and maintain performance by updating themselves with the real target views of past predictions. However, for real-time of OVSs, the revealing of past prediction targets cannot keep up with the new round of prediction, making existing models fail to achieve their promise. Thus, a reliable and timely indication of videos’ future views is critical for popularity prediction in OVSs. Recent studies show that the average watched percentage of a video’s content is correlated with the video’s views in the same period and is predictable through time series models. Inspired by these studies, we propose a model to infer videos’ popularity through prediction of the videos’ average watched percentage. Specially, through large-scale analysis of OVS data, we discover that a video age varying relation exists between the average watched percentage and the future views of videos. We then model the relation between the average watched percentage and the future views of videos through a video age-sensitive function, making the feature’s indication of future views reliable for different aged videos. Meanwhile, by restricting the temporal range of average watched percentage prediction, we are able to promptly update the feature’s prediction model, further stabilizing the performance of popularity prediction. We evaluate the proposed model on a real-world data set of iQIYI, one of the most popular OVS providers in China. The experimental results demonstrate the effectiveness of the proposed model.