Mining odd-length paths for link prediction in bipartite networks

Link prediction refers to predicting the possibility of a missing link in a network through network topology structures and/or node properties. However, traditional methods do not perform well in bipartite networks, in which the nodes are divided into two distinct sets, and the nodes of the same set cannot be connected. Predicting potential links in bipartite networks has many valuable applications in real life, such as compound-protein interaction prediction, gene-disease relationship identification, and item recommendation. Different from many prior efforts, this study tailors several link prediction methods for bipartite networks by mining specifically odd-length paths. Inspired by Katz, we first propose a new link prediction method, LPOP that naturally considers all odd-length paths between two nodes in a bipartite network. Then, we propose LPOPE, which improves LPOP by considering the similarities of the nodes of the same set along an odd-length path between two unconnected nodes. Experimental results on different networks indicate that our proposed link prediction methods predict potential links more accurately than the traditional methods. The average area under the curve (AUC) improvement rates of LPOPE or LPOP over the best baseline methods on real-world networks are 27.7%. Moreover, based on the results, it is recommended to use LPOPE if the average degree of a bipartite network is less than 5.0 since LPOPE additionally considers the node similarities on the paths with a length of three, and LPOP is recommended if a network is denser due to its simplicity.