A Comparative Experimental Study of Link Prediction Methods with Structural Information

Abstract

Link prediction is an important task to predict missing or future links in complex networks, social networks, knowledge graphs, etc. Since networks naturally have topological structures, a key issue is how to use structural information. Existing methods for link prediction can be categorized into two types: heuristic-based and learning-based. This paper compares these two types of methods and explores the factors affecting the performance. Experiments on five real-world datasets showed that the learning-based methods outperform the heuristic-based method, and their link prediction performance is affected by the size of node coverage. For learning-based methods, training time can be reduced by using smaller training set with enough node coverage.