With the rapid growth of social media, rumors are also spreading widely on social media and bring harm to people's daily life. Nowadays, information credibility evaluation has drawn attention from academic and industrial communities. Current methods mainly focus on feature engineering and achieve some success. However, feature engineering based methods require a lot of labor and cannot fully reveal the underlying relations among data. In our viewpoint, the key elements of user behaviors for evaluating credibility are concluded as "who", "what", "when", and "how". These existing methods cannot model the correlation among different key elements during the spreading of microblogs. In this paper, we propose a novel representation learning method, Information Credibility Evaluation (ICE), to learn representations of information credibility on social media. In ICE, latent representations are learnt for modeling user credibility, behavior types, temporal properties, and comment attitudes. The aggregation of these factors in the microblog spreading process yields the representation of a user's behavior, and the aggregation of these dynamic representations generates the credibility representation of an event spreading on social media. Moreover, a pairwise learning method is applied to maximize the credibility difference between rumors and non-rumors. To evaluate the performance of ICE, we conduct experiments on a Sina Weibo data set, and the experimental results show that our ICE model outperforms the state-of-the-art methods.