CST-Net: community-guided structural-temporal convolutional networks for popularity prediction

The ability to predict the popularity of online contents has important implications in a wide range of areas. The challenge of this problem comes from the inequality of the popularity of content and the numerous complex factors. Existing works fall into three main paradigms: feature-driven approaches, generative models, and methods based on deep learning, each with known strengths and limitations. In this article, we propose an end-to-end deep learning framework, called CST-Net, to combat the defects of existing methods. We first learn a low-dimensional embedding for each user based on historic interactions. Then, users are clustered into communities based on the learned user embeddings, and information cascades are represented as a series of episodes in the form of community interaction matrix. Afterwards, a convolutional architecture is applied to learn the representation of the entire information cascade. Finally, the extracted structural and temporal features are further combined to predict the incremental popularity. We validate the effectiveness of the proposed CST-Net by applying it on two different types of population-scale datasets, i.e., a microblogging dataset and an academic citation dataset. Experimental results demonstrate that the proposed CST-Net model consistently outperforms the existing competitive popularity prediction methods.