Line Graph Neural Networks for Link Weight Prediction

Link weight prediction is of great practical importance, since real-world networks are often weighted networks. Previous studies have mainly used shallow graph features for link weight prediction, which limits the prediction performance. In this paper, we propose a new link weight prediction algorithm, namely Line Graph Neural Networks for Link Weight Prediction (LGLWP), which learns deeper graph features through deep learning. In our algorithm, we first extract the enclosing subgraph around a target link, and then employ a weighted graph labeling algorithm to label the subgraph nodes. Next, we transform the subgraph into a line graph and apply the graph convolution neural networks to learn the node embedding in the line graph, which can represent the links in the original subgraph. Finally, the link feature vectors are put into a fully-connected neural network to predict the weight of the target link. Our algorithm directly obtain the feature vectors of the target links in the original graph, which is better than the previous methods that splice the node feature vectors for link weight prediction. Experiments results on six real datasets of various network sizes and types show that our algorithm has better prediction performance than the state-of-art methods, while it has fewer parameters and high training efficiency.