Predicting individual retweet behavior by user similarity: A multi-task learning approach

Abstract

Users read microblogs and retweet the most “interesting” tweets to their friends in online social networks. Predicting retweet behavior is extremely challenging due to various reasons. First, the most of existing approaches primarily discuss a global retweet predicting model, with a goal of finding a uniform model that fits all users, but ignore individual behavior. And while social influence plays an important role in information diffusion, this fact has been largely ignored in conventional research. In this paper, we adopt a “microeconomics” approach to a model, and predict the individual retweet behavior. We study relationships between users by considering social similarity, which reflects how a particular retweeting action affects both the originator and the receiver of the retweet. To address the individual and social challenges, we analyze the effect of social similarity on retweet behavior based on a real dataset. Moreover, we cast our predicting problem as a multi-task learning problem. Combining the social and individual understanding, we then propose a novel model for predicting individual retweet behavior. We conduct extensive experiments on a Weibo (http://weibo.com, the largest microblogging service in China) dataset to validate the effectiveness of the proposed model. Our results demonstrate the superior performance of the proposed model, compared with several alternative classification methods.