Enhancing graph convolutional networks with an efficient k-hop neighborhood approach

Abstract

Graph Convolutional Network (GCN) has emerged as a powerful model for network data analysis and representation since it effectively utilizes neighborhood information via message propagation. However, existing GCNs cannot efficiently utilize the deeply propagated k-hop neighborhood in-formation when k becomes large due to over-smoothing issues, which significantly constrains their capacity for representation learning to the deep structure of the graph. Motivated by this critical issue, this paper proposes an efficient k-hop Neighborhood Enhanced Graph Convolutional Network (𝑘NE-GCN) model is proposed in this paper, which is developed based on two-fold main ideas: a) deriving an efficient 𝑘-hop neighborhood enhancement scheme from the perspective of path-aware embedding, which enables each ego node to preserve the 𝑘-hop neighborhood information in polynomial form; and b) building a nearest-neighborhood constraint into the objective function for stressing one-hop neighborhood information, thus relieving the impacts by redundant nodes or noisy links. Empirical results from four benchmark datasets against 11 state-of-the-art models clearly illustrate the superior performance of the proposed kNE-GCN in missing link estimation in undirected weighted graphs.