Learning content–social influential features for influence analysis

We address how to measure the information propagation probability between users given certain contents. In sharp contrast to existing works that oversimplify the propagation model as predefined distributions, our approach fundamentally attempts to answer why users are influenced (e.g., by content or relations) and whether the corresponding influential features (e.g., hidden factors) can be inferred from the propagation in the entire network. In particular, we propose a novel method to deeply learn the unified feature representations for both user pair and content, where the homogeneous feature similarity can be used to estimate the propagation probability between users with given content. The features are dubbed content–social influential feature since we consider not only the content of the propagation information but also how it propagates over the social network. We design a fast asynchronous parallel algorithm for the feature learning. Through extensive experiments on a real-world social network with 53 million users and 838 million tweets, we show significantly improved performance as compared to other state-of-the-art methods on various social influence analysis tasks.