Network embedding for link prediction: The pitfall and improvement

Link prediction plays a significant role in various applications of complex networks. The existing link prediction methods can be divided into two categories: structural similarity algorithms in network domain and network embedding algorithms in the field of machine learning. However, few researchers focus on comparing these two categories of algorithms and exploring the intrinsic relationship between them. In this study, we systematically compare the two categories of algorithms and study the shortcomings of network embedding algorithms. The results indicate that network embedding algorithms have poor performance in short-path networks. Then, we explain the reasons for this phenomenon by computing the Euclidean distance distribution of node pairs after a given network has been embedded into a vector space. In the vector space of a short-path network, the distance distribution of existent and nonexistent links are often less distinguishable, which can sharply reduce the algorithmic performance. In contrast, structural similarity algorithms, which are not restricted by the distance function, can represent node similarity accurately in short-path networks. To address the above pitfall of network embedding, we propose a novel method for link prediction aiming to supplement network embedding algorithms with local structural information. The experimental results suggest that our proposed algorithm has significant performance improvement in many empirical networks, especially in short-path networks. AUC and Precision can be improved by 36.7%-94.4% and 53.2%-207.2%, respectively.