|
Abstract
|
Link prediction plays a vital role in understanding and modeling the dynamic behavior of complex networks such as social, biological, and citation networks. In real-world systems, many connections are missing, hidden, or yet to appear in the future. Accurately predicting these potential links helps improve tasks like friend recommendations, knowledge graph completion, biological interaction discovery, and network evolution analysis. Traditional link prediction methods mainly rely on topological similarity (for example, Common Neighbors or Adamic-Adar), which only capture static structural information. However, in real networks, link formation is also strongly influenced by node influence that reveals the ability of some nodes to spread information or attract new connections. Ignoring this factor leads to inaccurate or biased predictions. On the other hand, deep learning models such as Graph Convolutional Networks (GCNs) have recently shown high performance in learning graph representations, but most of them do not explicitly include node influence or adaptive similarity measures. Therefore, there is a clear need to design a hybrid link prediction framework that can effectively combine both influence-aware features and structural similarity within a unified deep learning model. This research is important because it aims to fill this gap by integrating social influence analysis with graph representation learning. The results can contribute to better network understanding, higher prediction accuracy, and broader applications in social media analysis, recommendation systems, and biological network inference.
|