An End-to-End Structure Aware Graph Convolutional Network for Modeling Multi-relational Data

Abstract

Low-dimensional embeddings of entities and relations in large scale knowledge graphs have been proved extremely beneficial in variety of downstream tasks, e.g. entity classification and knowledge graph completion. Most of existing approaches incorporate both textual information and relation paths of triple facts for knowledge graph representation. However, they ignore rich structural information in a knowledge graph, i.e., connectivity patterns in neighboring entities and relations around a given entity. In this work, we propose a novel knowledge representation model, denoted Structure Aware Graph Convolutional Network (SAGCN), which leverages structural information for modeling the highly multi-relational data characteristic of realistic knowledge graphs. Specifically, we sample multi-hop neighboring entities and relations of a given entity as its local graph, which depicts the neighborhood topology structure. To encode features from the local graph, we introduce localized graph convolutions as a neighborhood structure encoder to generate embeddings. We further design distinct decoders for entity classification and knowledge graph completion. The proposed approach are evaluated on three public datasets and substantially outperforms state-of-the-arts.