Effective cross-catalogue analysis is crucial in astronomical research; however, existing systems are hindered by sparse relationships between catalogues. To address this challenge, we propose a graph neural network-based approach for recommending astronomical catalogues. We begin by constructing a multi-relational dataset with catalogues in VizieR, integrating various relationships and catalogue attributes. Then a method for encoding textual information using pre-trained embedding models is introduced, coupled with a two-stage training strategy designed to enhance link prediction performance. Our experimental results demonstrate that the proposed method outperforms existing models. Additionally, a large language model is utilized to perform a secondary assessment of the top-k recommended results, providing users with the rationale behind these evaluations, enhancing both the accuracy and interpretability of the recommendations.