COMPutational Analysis of Social System Lab

Link prediction is a fundamental task in network science, with applications ranging from recommender systems to inferring biological interactions. Most link prediction methods are evaluated under the assumption that missing edges are uniformly distributed throughout the network. However, in real-world scenarios, missing edges often follow non-uniform patterns that depend on node attributes or network structure.

In this study, we investigate how different missing-edge patterns affect the performance of common link prediction algorithms. We analyze various real-world networks and demonstrate that prediction accuracy can vary significantly depending on the underlying pattern of missing edges. Our findings suggest that the choice of link prediction method should be informed by the specific characteristics of the missing data in the network under study.

We propose a framework for evaluating link prediction methods under different missing-edge scenarios and provide recommendations for selecting appropriate algorithms based on the expected pattern of missing edges. This work contributes to a more realistic understanding of link prediction performance in practical applications and highlights the importance of considering the non-uniform nature of missing data in network analysis.