GraphAIR: Graph representation learning with neighborhood aggregation and interaction

Graph representation learning is of paramount importance for a variety graph analytical tasks, ranging from node classification to community detection. Recently, convolutional networks (GCNs) have been successfully applied learning. These GCNs generate by aggregating features the neighborhoods, which follows “neighborhood aggregation” scheme. In spite having achieved promising performance on various existing GCN-based models difficulty in well capturing complicated non-linearity data. this paper, we first theoretically prove that coefficients neighborhood interacting terms are relatively small current models, explains why barely outperforms linear models. Then, order better capture data, present novel GraphAIR framework interaction addition aggregation. Comprehensive experiments conducted benchmark tasks including and link prediction using public datasets demonstrate effectiveness proposed method.