A Novel Time Series Approach for Predicting the Long-Term Popularity of Online Videos

Abstract

Predicting the video popularity is an essential part of fast growing online media services. It is beneficial to an array of domains, from targeted advertising, personalized recommendation, to traffic load optimization. However, popularity prediction is a challenge problem due to the uncertainty of information cascade. In this paper, we treat the popularity of online videos as time series over the given periods and propose a novel time series model for popularity prediction. The proposed model is based on the correlation between early and future popularity series. Instead of inferring the precise view counts for a video, this paper focuses on accurately identifying the most popular videos based on the predicted popularity, because it is of the most interest to service providers. Experimental result on real world data have demonstrated that the proposed model outperforms several existing popularity prediction models.