We developed user models of knowledge exploration in a social tagging system to test the expertise rankings generated by a link-structure method and a semantic-structure method. The link-structure method assumed a referential definition of expertise, in which experts were users who tagged resources that were frequently tagged by other experts; the semantic-structure method assumed a representational definition of expertise, in which experts were users who had better knowledge of a particular domain and were better at assigning distinctive tags associated with certain domain-specific resources. Simulations results showed that the two methods of expert identification, although based on different assumptions, were in general consistent but did show significant differences. As expected, the link-structure method was better at facilitating exploration of popular "hot" topics than the semantic-structure method. However, the semantic-structure method was better at guiding users to find less popular "cold" topics than the link-structure method. Resources tagged by domain experts could contain cold topics that were associated with high quality tags, but these resources were less likely highlighted by the link-structure method. We argue that to facilitate knowledge exploration in social tagging systems, it is important to keep a good balance between helping user to follow hot topics and to discover cold topics by including expertise rankings generated by both link and semantic structures.