With the surge of academic papers, it has become a common practice to recommend papers based on authors’ research interests. Existing methods focus on leveraging author–paper research interactions to mine authors’ research interests with coauthorship networks. However, sparse research interactions would pose a huge challenge to distinguish research interests of authors. Fortunately, inter-dependent knowledge across papers provides rich potential heterogeneous connections for author–paper interactions, offering much insights for learning authors’ research interests. Therefore, we propose a meta-relation–guided knowledge coupling approach for paper recommendation. Specifically, we construct a meta-relation–guided heterogeneous graph architecture to depict the numerous inter-dependencies among authors and papers, thereby exploring complex author–paper interactions. First, a meta-relation–aware heterogeneous graph encoder is developed to extract relational structure which maintains the relation-specific representation of authors’ research interest and papers’ research relatedness. Then, a cross-meta-path attention network is designed to aggregate the characteristics of different meta-relations and obtain research features of authors and papers. Finally, a self-supervised data augmentation architecture is constructed to mine and preserve local and global graph structure information, acquiring papers with high relevance to author’s research interests through training loss. Numerous experiments are conducted on two real academic datasets, effectively demonstrating the superiority of our proposed model and validating its effectiveness in paper recommendation.