Mining node attributes for link prediction with a non-negative matrix factorization-based approach

Link prediction determines if there is an edge between two unconnected nodes in a complex network using known information, such as network topology and/or node semantic attributes. However, existing link prediction methods primarily rely on topological structure of two unconnected nodes. In recent years, an increasing number of networks have provided node semantic attributes that are important to link prediction. In general, similar nodes are more likely to be linked. Based on non-negative matrix factorization (NMF), this study proposes a novel link prediction method, NMFLP, to integrate network topology and node attributes. NMFLP can be employed for two types of networks: for attribute networks with node semantic attributes, it can directly use node semantic attributes and predict possible links by flexibly integrating network topology and node semantic attributes; for networks without attributes, it uses the topological structure information of a node as its attribute. To evaluate the performance of NMFLP, this study compares it with both classical and state-of-the-art baseline methods on different networks. On the real-world and synthetic networks with semantic attributes, the average improvement rates of NMFLP are 7.93%–57.00% and 10.18%–30.43%, respectively, in terms of area under the curve (AUC), whereas on the real-world and synthetic networks without semantic attributes, the average AUC improvement rates of NMFLP are 11.29%–71.69% and 3.54%–187.29%, respectively. NMFLP also exhibits good convergence performance during the matrix factorization process of NMF.