Large-scale text categorisation in social environments, characterised by the high dimensionality of feature spaces, is one of the most relevant problems in machine learning and data mining nowadays. Short-texts, which are posted at unprecedented rates, accentuate both the importance of learning tasks and the challenges posed by such large feature space. A collection of social media short-texts does not only provide textual information but also topological information given by the relationships between posts and their authors. The linked nature of social data causes new complementary data dimensions to be added to the feature space, which, at the same time, becomes sparser. Additionally, in the context of social media, posts usually arrive simultaneously in streams, which hinders the deployment of efficient traditional feature selection techniques that assume a feature space fully known in advance. Hence, efficient and scalable online feature selection becomes an important requirement in numerous large-scale social applications. This work presents an online feature selection technique for high-dimensional data based on the integration of two information sources, social and content-based, for the real-time classification of short-text streams coming from social media. It focuses on discovering implicit relations amongst new posts, already known ones and their corresponding authors to identify groups of socially related posts. Then, each discovered group is represented by a set of non-redundant and relevant textual features. Finally, such features are used to train different learning models for classifying newly arriving posts. Extensive experiments conducted on real-world short-texts demonstrate that the proposed approach helps to improve classification results when compared to state-of-the-art and traditional online feature selection techniques.