Link Prediction in Temporal Networks Using Graph-Regularized Concept Factorization

The goal of temporal link prediction is to predict future networks based on existing network information. Among the various techniques for link prediction, Non-negative Matrix Factorization (NMF) is widely utilized. However, traditional NMF methods fail to effectively incorporate the diverse characteristics of networks. Concept factorization (CF), as a variant of NMF, can address this issue effectively. Therefore, this paper applies concept factorization to the link prediction problem and proposes a graphregularized concept factorization model (GrCF). To avoid the loss of temporal dependencies caused by network collapsing, the model integrates network data from time slice 1 to time slice $t-1$ as regularization to capture information within the temporal slices. The approach combines the second-order similarity matrix and adjacency matrix as input to extract higher-order network information. Additionally, a smoothing regularization term is introduced to alleviate overfitting. Furthermore, an efficient alternating iterative method with convergence guarantees is proposed to minimize the model. The model is tested on several synthetic and real-world networks, with results demonstrating its superiority over other methods.