We describe a method to derive design models for hypermedia interfaces from the bottom up. Firstly, we compile a list of hypermedia interface features which we classify according to the category of functions they fulfill. We then describe an experiment in which candidate designs for low-level interface features were designed and tested for recognizability. In the experiment, icons for each of 61 hypermedia concepts were generated and then judged. Finally, we outline and illustrate a model induction phase in which low-level features are combined into an overall interface model, via "micro-models" that take account of the types of icons that worked best for each class of interface feature. We suggest that, at least for hypermedia systems, a bottom-up approach to interface design based on the functions of low-level features is preferable to the dominant, top-down approach based around one or more metaphors. Copyright 1993, 1999 Academic Press