|
Abstract:
|
Social media has become an important platform for people to share news and information about major events. In this paper, we examine whether we can make predictions at the early stage of the diffusion about users' future influence in the context of this event. We intend to identify individuals who will be important sources of information during the event, using the aggregate size of the retweet cascades of the user as a measure of influence. In addition to a few commonly used user features that have been applied in previous research. We have also extracted new features derived from the user's recent historical retweet network. Given limited information, we extract and employ all of these features to perform an "online" prediction about users' future cascades sizes using several state-of-art machine learning methods, including deep learning. We show the addition of new features derived from the diffusion network give better results. We give some suggestions for practical implementation. We argue that our framework is currently the best available for early prediction during urgent event diffusion and is also highly compatible with the limitations of the Twitter streaming AP
|
