An improved network embedding method with multi-level closeness on link prediction

Abstract

Network representation learning provides an important tool to link prediction in complex networks. Many existing methods consider the random walk within the direct neighbors of the nodes, however, ignore the closeness level between nodes. In this article, we propose a novel network embedding method by considering the closeness of three different levels, i.e., the close, median and faraway relationships. The close relationship is modeled by a natural nearest neighbor, the median relationship is referred to as the direct neighbor, and the faraway relationship is simulated by a role discovery. Diversified learning can better capture the node feature, and therefore helps improving link prediction. Experimental results show that the proposed method outperforms nine baseline methods, by testing them on six real datasets. The closenesses of the three levels are found to impact differently on the networks. In general, the direct neighbor closeness has a great impact, however, for the network with specific characteristics, other closenesses may be more important, e.g., the role neighbor closeness is important in the economic network.