A structure distinguishable graph attention network for knowledge base completion

A knowledge graph is a collection of triples, often represented in the form of “subject,” “relation,” “object.” The task of knowledge graph completion (KGC) is to automatically predict missing links by reasoning over the information already present in the knowledge graph. Recent popularization of graph neural networks has also been spread to KGC. Typical techniques like SACN achieve dramatic achievements and beat previous state-of-the-art. However, those models still lack the ability to distinguish different local structures within a graph, which leads to the over smoothing problem. In this work, we propose SD-GAT, a graph attention network with a structure-distinguishable neighborhood aggregation scheme, which models the injective function to aggregate information from the neighborhood. The model is constituted of two modules. The encoder is a graph attention network that improved with our neighborhood aggregation scheme, which could be applied for a more distinct representation of entities and relations. The decoder is a convolutional neural network using 3×3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document} convolution filters. Our empirical research provides an effective solution to increase the discriminative power of graph attention networks, and we show significant improvement of the proposed SD-GAT compared to the state-of-the-art methods on standard FB15K-237 and WN18RR datasets.