Text contents generated by social networking platforms tend to be short. The problem of authorship attribution on short texts is to determine the author of a given collection of short posts, which is more challenging than that on long texts. Considering the textual characteristics of sparsity and using informal terms, we propose a method of learning text representations using a mixture of words and character n-grams, as input to the architecture of deep neural networks. In this way we make full use of user mentions and topic mentions in posts. We also focus on the textual implicit characteristics and incorporate ten latent posting styles into the models. Our experimental evaluations on tweets show a significant improvement over baselines. We achieve a best accuracy of 83.6%, which is 7.5% improvement over the state-of-the-art. Further experiments with increasing number of authors also demonstrate the superiority of our models.