Information Retrieval provides numerous challenges for Machine Learning: The sheer size of current document collections creates a high demand for tools that aid their navigation but also pushes the boundaries in terms of required algorithmic efficiency. At the same time, the only "ground truth" available is the relevance as perceived by the user, but user preferences may be a moving target. In this subject area, we are investigating different approaches to exploit hidden structures in user feedback and document contents. Furthermore, many structures in online information systems can be represented as graphs and can be mined for generating user profiles. Here, we are currently analysing datasets created by social web applications using hypergraphs for data representation. Applications include community detection, i.e. the identification of particularly well-connected subgroups, and the clustering of documents using tagging data.<br /> <br />Acknowledgements: Research was funded by DFG and Technische Universität Berlin.