Temporal Convolutional Networks For Popularity Prediction Of Messages On Social Medias

Predicting the popularity of messages on social medias is an important problem that draws wide attention. The temporal information is the most effective one for predicting future popularity and has been widely used. Existing methods either extract various hand-crafted temporal features or utilize point process to modeling the temporal sequence. Unfortunately, the performance of the feature-based methods heavily depends on the quality of the heuristically hand-crafted features while the point process methods fail to characterize the longer observed sequence. To solve the problems mentioned above, in this paper, we propose to utilize Temporal Convolutional Networks (TCNs) for predicting the popularity of messages on social media. Specifically, TCN can automatically adopt the scales of observed time sequence without manual prior knowledge. Meanwhile, TCN can perform well with long sequences with its longer effective memory. The experimental results indicate that TCN outperforms all the baselines, including both feature-based and point-process-based methods.