Graph-based filtering (GF) methods are suitable for addressing paper/citation recommendation problems. However, these methods still produce biased rankings due to multi-topic, similar, and related topics. Moreover, a scholarly dataset is expanding into big data, causing interesting issues in citation recommendation, especially information overload. Hence, we introduce a topic community selection model to take text and graph analysis into account. The model consists of four stages: transforming datasets into citation networks, detecting communities, identifying topics in communities, and selecting topic communities. The proposed method can significantly limit the search space but still produces good recall accuracy. The experiment shows that the community selection relating to specific topical queries produces five selected sub-graphs (communities) constituting only 18% of the total dataset volume while having 97% recommended paper candidates of the ground truth test. In future research, paper recommendations using GF can be carried out sufficiently on the sub-graph rather than the whole graph.