Link Analysis in Complex Networks by Structural Methods and Machine Learning

The analysis of connections in complex networks has recently aroused much interest in systems biology, which includes the development of medicine and the description of protein interaction, as well as in concept processing, where semantic association networks are of importance. Such investigation requires addressing the link prediction problem, which implies restoring the missing connections. Most structural methods for solving this problem have been originally proposed for social networks, and might not provide necessary performance in some networks of other nature. In this paper, we provide the comparison of structural methods and representations learning techniques in terms of link prediction performance in the formal classification task for various biological networks and semantic network of English word associations. Our results show that representation learning methods remain more resistant to the information leak in the considered networks, and can noticeably outperform basic structural indices in terms of link prediction. Regarding the semantic network, we manage to reveal the impact of association strength on restoring the structure of network.