Abstract

Graph Neural Networks (GNNs) have been widely used for various learningtasks, ranging from node classification to link prediction. They havedemonstrated excellent performance in multiple domains involvinggraph-structured data. However, an important category of learning tasks, namelylink weight prediction, has received less emphasis due to its increasedcomplexity compared to binary link classification. Link weight predictionbecomes even more challenging when considering multilayer networks, where nodescan be interconnected across multiple layers. To address these challenges, wepropose a new method named Multiplex Spatial Graph Convolution Network (MSGCN),which spatially embeds information across multiple layers to predict interlayerlink weights. The MSGCN model generalizes spatial graph convolution tomultiplex networks and captures the geometric structure of nodes acrossmultiple layers. Extensive experiments using data with known interlayer linkinformation show that the MSGCN model has robust, accurate, and generalizablelink weight prediction performance across a wide variety of multiplex networkstructures.