Effect of second-order network structure on link prediction

Abstract

Small-degree nodes widely exist in real networks, causing the difficulty in link prediction for them due to the lack of information. The clustering information benefits the link prediction by introducing the network inner structure, however, the commonly discussed first-order clustering information is still insufficient for the link prediction of the small-degree nodes. In this article, we introduce the second-order network structure to complement information for the small-degree nodes. A general link prediction approach is proposed by incorporating the second-order clustering coefficient, and is employed to improve eight baseline algorithms. Experimental results show that all the baseline algorithms are remarkably improved. Compared with three advantageous similarity-based and two learning-based algorithms, an improved common neighbor method also shows an advantage in most cases. Further, an information gain between the first- and the second-order network structure is investigated, and the second-order network structure is found to also contain abundant information, which provides a possible understanding to the proposed approach. Our work may shed a new light on how network structure affects link prediction.