Abstract
Link prediction, which aims to estimate missing or future connections in networks, is an important problem with a wide range of applications. Traditional similarity-based link prediction methods exploit local structural features but fail to capture community structures. This paper proposes a weighted link prediction method that incorporates community detection algorithms for computing the proposed methods. Four real-world weighted networks from different domains are analyzed using three established community detection algorithms - Louvain, Girvan-Newman, and ALPA. The identified community structures are then utilized to augment five traditional weighted link prediction methods - WCN, WPA, WAA, WJC, and WRA. Experimental results on the four networks show that the proposed community-informed link prediction approach significantly outperforms the baseline methods, achieving improvements in AUC ranging from 0.32–13.62%. Further analysis indicates that the performance boost depends on the network topology, community structure, and properties of different prediction algorithms. This work demonstrates the importance of leveraging global network structures beyond local features for more accurate link prediction, especially in sparse and scale-free networks. The proposed methods can help advance and apply link prediction across complex networked systems.