Dynamic Heterogeneous Graph Learning: An Adaptive Research Academic Network

Accurately assessing the impact of academic achievements holds significant importance for scholars in their literature review process and for the retrieval and recommendation of scientific research databases. Predicting this impact presents a formidable challenge. Furthermore, scientific research and academic networks exhibit dynamic evolutionary characteristics. Over time, not only does the semantic information of keywords, journals, and other nodes undergo transformation, but the strength of connections between distinct nodes also experiences fluctuations. In recent years, with the advancement of deep learning technologies, particularly the introduction of recurrent neural networks, graph neural networks, and related architectures, a powerful tool for data representation learning has emerged. This paper adopts a dynamic graph representation learning perspective, aiming to adaptively derive vectorized representations for each node through the design of a trainable neural network, ultimately predicting the influence of academic achievements. The study centers on publicly available scientific research and academic networks like APS and AMiner, employing the citation count of papers as the evaluation metric for influence. Specifically, this research will formulate a trainable neural network from a data-driven standpoint to dynamically capture semantic information and evolutionary trends within dynamic graph structures, subsequently generating corresponding vector representations for individual nodes. Upon acquiring the semantic representation of the article, the future citation count can be forecasted through the design of a straightforward mapping function, such as a multi-layer perceptron. Furthermore, through the analysis of node representations (including authors, journals, etc.), it is possible to uncover and explore the evolutionary patterns of individuals and groups within the scientific research academic network.