基 于 图 神 经 网 络 的 学 术 论 文 被 引 频 次 预 测 研 究

［Purpose/significance］ Predicting the citation frequency of academic papers is a crucial aspect of evaluating the quality and impact of scholarly work. ［Method/process］ This paper introduces Graph Neural Networks （GNN） to construct a heterogeneous network encompassing papers，authors，and journals based on the S2ORC dataset. By leveraging the sampling and aggregation mechanisms of the GraphSAGE model，we enhance prediction accuracy and progressively incorporate journal and author sub-networks into the citation network to explore potential heterogeneities in predictions across different disciplinary fields. ［Result/conclusion］ The GraphSAGE model outperforms GCN and GAT models in high average degree networks but performs less effectively in domains or network structures where global citation relationships are more deeply dependent，where GCN shows better performance. When the author subnetwork is included alone， there is an improvement in prediction performance in the fields of arts and law，while performance declines in other disciplines；when both journal and author nodes are introduced simultaneously，only the fields of arts and law exhibit enhanced predictive performance.