Influence prediction plays a crucial role in the academic community. The amount of scholars' influence determines whether their work will be accepted by others. Most existing research focuses on predicting one paper's citation count after a period or identifying the most influential papers among the massive candidates, without concentrating on an individual paper's negative or positive impact on its authors. Thus, this study aims to formulate the prediction problem to identify whether one paper can increase scholars' influence or not, which can provide feedback to the authors before they publish their papers. First, we presented the self-adapted ACC (Average Annual Citation Counts) metric to measure authors' impact yearly based on their annual published papers, paper citation counts, and contributions in each paper. Then, we proposed the RD-GAT (Reference-Depth Graph Attention Network) model to integrate heterogeneous graph information from different depth of references by assigning attention coefficients on them. Experiments on AMiner dataset demonstrated that the proposed ACC metrics could represent the authors influence effectively, and the RD-GAT model is more efficiently on the academic citation network, and have stronger robustness against the overfitting problem compared with the baseline models. By applying the framework in this work, scholars can identify whether their papers can improve their influence in the future.

翻译：影响力预测在学术界发挥着至关重要的作用。学者影响力的强弱决定了其研究成果能否被他人接受。现有研究大多聚焦于预测单篇论文在未来一段时期的被引频次，或从海量候选论文中识别最具影响力的文献，而较少关注单篇论文对其作者产生的积极或消极影响。因此，本研究旨在构建一种预测框架，以判别单篇论文是否能够提升学者的学术影响力，从而为作者在发表论文前提供决策反馈。首先，我们提出了自适应ACC（年均被引频次）指标，该指标基于学者每年发表的论文、论文被引频次及其在每篇论文中的贡献度，实现对其年度影响力的量化评估。随后，我们提出RD-GAT（参考文献深度图注意力网络）模型，通过为不同引用深度的参考文献分配注意力系数，实现对异质图信息的融合。在AMiner数据集上的实验表明：所提出的ACC指标能有效表征学者影响力；与基线模型相比，RD-GAT模型在学术引用网络中具有更高预测效率，并对过拟合问题展现出更强的鲁棒性。通过应用本研究的框架，学者可预判其论文是否有助于提升未来的学术影响力。