<![CDATA[Co-authorship prediction is important in academic network analysis due to it helps to understand patterns of scientific collaboration and supports collaboration recommendation systems. Topology-based approaches, such as connectivity metrics and node distance, have been widely used to model new relationships in networks. However, these approaches often overlook relevant author attributes, such as reputation and productivity. This study develops a co-authorship prediction model by combining a Graph Attention Network (GAT) and a Random Forest. GAT is used to extract topological features from the co-authorship graph, while Random Forest leverages additional attributes such as h-index and the number of publications to improve prediction accuracy. Experiments were conducted on a co-authorship dataset comprising over 10,000 authors and 50,000 publications. The results show that GAT achieved 85% accuracy, while Random Forest reached 80%. The combination of the two yielded 90% accuracy and a higher F1-score, indicating a better balance between precision and recall. The combined model also proved more accurate in predicting collaborations involving highly productive authors. These findings suggest that a hybrid approach can more comprehensively capture the dynamics of academic collaboration and may serve as a foundation for developing more effective collaboration prediction systems in the future.]]>
