Truncated Graph-Regularized Low Rank Representation for Link Prediction

Link prediction, whose primitive aim lies in remodeling or inferring link formations in complex networks, has been accepted as a fundamental study in understanding interactions between specific node pairs. To overcome the shortcomings of sparsity and intricacy in networks, an iterative method is proposed to comprehensively describe links in both local and global perspectives. From defining a truncated similarity matrix, local inherited properties in the network are maintained. With the refined similarity, a graph-regularized low-rank representation is provided to simultaneously preserve local and global structure information. Then, the representation is optimized to accurately predict link interactions in the network. Compared with the state-of-arts on real-world networks, the competitive experimental results demonstrate that our method is capable of effectively delineating interactions in multiple networks.