On modeling and predicting popularity dynamics via integrating generative model and rich features

Abstract Understanding the mechanisms governing how an online message acquires more popularity than another, modeling how it gains popularity dynamically, and determining the method for predicting its dynamics popularity are of tremendous interest to related decision support systems. However, one major limitation of existing generative dynamics models is that the learning parameters are difficult to interpret and it is unclear whether it can be generalized for other messages, as they are trained for different messages independently and the feature data-based connections between messages are ignored. To alleviate the defects, we first perform experiments on real-world data from Sina Weibo to identify the general correlation between the dynamics model and rich features of online messages. Consequently, we present a novel feature-regularized dynamics model based on reinforced Poisson process (FRRPP), which regulates the parameter learning of popularity dynamics by integrating a feature regression term to capture the revealed correlation across online posts. Specifically, in addition to the objective of the maximum likelihood function, we assume that the competitiveness parameter of the different posts can be predicted by rich features, to enhance the explicability and generality of the point-process model and learn the dynamics process of different posts together. The proposed model is then evaluated on two real Sina Weibo datasets, and conclusive experimental results indicate that the proposed model achieves a remarkable improvement over baseline methods in terms of MAPE and Accuracy with various settings, which further verifies our findings about how to improve the generality of popularity dynamics modeling and prediction.