Abstract: Cancer research collaboration recommendation can be naturally formulated as a link prediction problem on continuous-time dynamic graphs. However, existing temporal graph models are primarily designed for general scenarios and struggle to incorporate domain-specific semantic information. To address this limitation, an adaptive temporal graph network model, termed AdaTGN-CR, is proposed for research collaboration recommendation. Specifically, an expert-driven adaptive neighbor sampling mechanism is introduced to dynamically adjust the neighborhood size and neighbor importance weights according to expert status. In addition, an expert feature modeling method is developed to quantify academic activity, collaboration intensity, and expertise level. Furthermore, a layer-normalized gated recurrent unit (GRU) is employed to enhance the stability of temporal memory updates and capture long-term collaboration patterns. Experimental results on the cancer research dataset CATCD demonstrate that the proposed model achieves average precision (AP) of 93.99% and93.81% under transductive and inductive settings, respectively, consistently outperforming state-of-the-art baseline models. These results indicate that integrating domain semantic information with adaptive temporal graph modeling significantly enhances the performance of research collaboration recommendation.

Key words: dynamic graph neural networks, research collaboration recommendation, expert features, temporal graph networks, adaptive neighbor sampling, link prediction