Exploring an edge convolution and normalization based approach for link prediction in complex networks

Abstract

Link prediction in complex networks is to discover hidden or to-be-generated links between network nodes. Most of the mainstream graph neural network (GNN) based link prediction methods mainly focus on the representation learning of nodes, and are prone to over-smoothing problem. This paper dedicates to the representation learning of links, and designs an edge convolution operation so as to realize the link representation learning. Besides, we propose an normalization strategy for the learned link representation, for the purpose of alleviating the over-smoothing problem of edge convolution based link prediction model, when constructing the link prediction graph neural network EdgeConvNorm with stacking edge convolution manipulations. Lastly, we employ a binary classifier sigmod on the Hadamard product of two nodes representation parsed from the final learned link representation. The EdgeConvNorm can also be employed as a baseline, and extensive experiments on real-world benchmark complex networks validate that EdgeConvNorm not only alleviates the over-smoothing problem, but also has advantages over representative baselines.