Gated Relational Graph Neural Network for Semi-supervised Learning on Knowledge Graphs

Entity classification is an important task for knowledge graph (KG) completion and is also crucial in many upper-level applications. Traditional methods use unsupervised representation learning to embed entities and relations into a continuous low-dimensional space, and then use the embeddings in downstream tasks. Recent years, Graph Neural Networks (GNNs) have been gaining growing interest, among which Graph Convolutional Network (GCN) is widely used in semi-supervised tasks due to its excellent capability of aggregating neighborhood features. However, GCN lacks the ability to deal with edge features, which is essential in KGs. In this paper, we propose Gated Relational Graph Neural Network (GRGNN) targeted on entity classification problem in KGs. More specifically, we apply the idea of TransE to incorporate features of entities and relations, and introduce gate mechanism to leverage hidden states of current node and its neighbors. Our method achieves state-of-the-art performance compared with other methods in FB15K and DB10K datasets.