Leveraging adaptive peeking window to improve self-exciting point process model for popularity prediction

Predicting the popularity of online content is an important issue. The mainstream method is to model the cumulative growth of the popularity as a temporal point process and to make predictions based on the observed initial period of information cascade. The peeking window, which will be taken into consideration in making predictions, is vitally important for the accuracy of predictions. However, the existing studies only generated hypotheses about the initial burst and maintained a consistent size of the peeking window for all content. The limited accuracy of previous approaches raises a fundamental question, i.e., How can we obtain the most effective part of the history to make an accurate prediction? In this paper, we identified the existence of a strong correlation between the peeking window and the temporal dynamic of the instantaneous relative attractiveness for a given online content. An investigation was conducted to explore the adaptive peeking window, which was used in a selfexciting point process model to predict eventual future popularity. Empirical studies on a Twitter dataset demonstrated that the proposed method significantly outperformed existing approaches.