In applications of formal concept analysis to real-world data, it is often necessary to model a reduced set of attributes to keep the resulting concept lattices from growing unmanageably big. If the results of the modeling are to be used by humans, e.g. in search engines, then it is important that the similarity assessment matches human expectations. We therefore investigated experimentally if the set-theoretic reformulation of Tversky's contrast model by Geist, Lengnink and Wille provides such a match. Predicted comparability and its direction was reflected in the human data. However, the model rated a much larger proportion of pairs as incomparable than human participants did, indicating a need for a refined similarity model.
展开▼
