Exploring Popularity Predictability of Online Videos With Fourier Transform

Abstract

The prediction of video popularity is of significant importance to online video service providers in terms of resource provisioning, online advertisement, and video recommendation. Traditional approaches normally utilize videos' historical popularity traces to make such a prediction. However, it is still uncertain whether the future popularity of a video is sure to be associated with its past popularity. In this paper, we explore the problem of video popularity predictability by analyzing videos' view count traces in the frequency domain with Fourier transform. We observe that sharp turns (e.g., peaks and valleys) of view count traces cause the inaccuracy in popularity prediction, which can be seized and quantified by high-frequency components in the frequency domain. Based on the ratio of high-frequency energy, videos can be classified as the fluctuating group, which is hard for prediction, and the smooth group, which is friendly for prediction. The result is further verified via experiments with state-of-the-art predictive algorithms. Inspired by our findings, we propose a strategy to improve prediction performance by removing out-of-date traces before each sharp turn because it is highly possible that the popularity evolution trend has been altered at each sharp turn. To the end, we compare the prediction issue between videos and microblogs. Surprisingly, most microblog traces are smooth. We conjecture that video providers' recommendation and promotion strategies are prone to causing sharp turns in view count traces. In contrast, there is no such initiative counterpart on microblog platforms changing trace evolution of microblogs frequently.

Keywords

• Popularity predictions
• Fourier transform
• video classification