Social Tagging Systems (STSs), allowing users to annotate online resources with freely chosen key words, are an essential type of application in Web 2.0. Recommendation in STSs can prevent information overload and support users to locate relevant items for interaction. This article applies a Topic-Graph Based Recommendation approach. First, we discover semantics behind tags through topic inferencing with Latent Dirichlet Allocation (LDA). Second, we conduct Graph-Based Recommendation for tags and users. The approach is applied on a real-word representative data sample collected from the Academic Social Networking Site ResearchGate. The widely used Co-occurrence Based Graph Recommendation is implemented as a baseline approach. Our preliminary human evaluation shows that the Topic-Graph Based Recommendation can complement to the Co-occurrence baseline to provide more reliable results. Future studies are provided on leveraging future features and information for recommendation from researcher-generated social media data on a large scale.