Unstructured textual data from online profiles is often used in conjunction with other user metadata to mine, in a supervised fashion, the latent demographic attributes of social media users (e.g. age, gender, occupation). Supervised methods, however, require labeled training data, which are often expensive to generate, and thus it would be attractive to re-use models across different domains and groups, i.e. training on a labeled dataset in order to mine the same latent attributes in those datasets for which training labels are missing. However, online conversations are often influenced by a myriad of topics and other factors, such as external events, and thus not all the features generated from this kind of data may perform well in a cross-domain setting. Here we study which of the features commonly found in public user profiles are portable across domains. As benchmark we focus on the very common task of detecting the gender of Twitter users from their public profile information — tweets, screen name, and profile picture. Our approach, based on a boosted stacked classifier, outperforms the state of the art in the task. Using data from two very different samples of Twitter users — one drawn from the public random stream and one about a recent social movement — we show that screen name and profile picture generalize across domains well, while text does not. Social media platforms have become attractive sources of data for computational approaches to social modeling, mainly due to their rapid growth and for the surprising ability to offer insight into real-world phenomena. Cross-domain user mining methods can help computational social science research by providing a richer and more accurate context to social phenomena.