Graph Convolutional Networks Based on 2D Convolution Operators

Most of the existing knowledge graphs suffer from incompleteness, which can predict the missing links based on known facts. A popular solution is to update the embedding representation of the central node by aggregating information from neighboring nodes and making inferences. CompGCN, a novel approach, can learn embeddings of nodes and relations simultaneously, which performs three composed operations (subtraction, multiplication, and circular-correlation) on neighboring nodes and corresponding relations. However, these three operations are too simplistic, do not consider the interaction of embeddings between nodes and relations, and capture limited semantic information. In this paper, we design a new model, Conv2DGCN, that combines GCN and 2D convolution. Conv2DGCN obtains rich feature interactions through 2D convolution, which allows nodes to aggregate more information. In addition, we use two methods for increasing feature interactions, namely “checkered” feature reshaping and circular convolution. Further, the MRR scores of Conv2DGCN are 6.8% and 7.5% better than CompGCN on FB15k-237 and WN18RR datasets, respectively.