Knowledge graph representation learning: A comprehensive and experimental overview

Abstract

Knowledge graph embedding (KGE) is a hot topic in the field of Knowledge graphs (KG). It aims to transform KG entities and relations into vector representations, facilitating their manipulation in various application tasks and real-world scenarios. So far, numerous models have been developed in KGE to perform KG embedding. However, several challenges must be addressed when designing effective KGE models. The most discussed challenges in the literature include scalability (KGs contain millions of entities and relations), incompleteness (missing links), the complexity of relations (symmetries, inversion, composition, etc.), and the sparsity of some entities and relations. The purpose of this paper is to provide a comprehensive overview of KGE models. We begin with a theoretical analysis and comparison of the existing methods proposed so far for generating KGE, which we have classified into four categories. We then conducted experiments using four benchmark datasets to compare the efficacy, efficiency, inductiveness, the electricity and the CO2 emission of five state-of-the-art methods in the link prediction task, providing a comprehensive analysis of the most commonly used benchmarks in the literature.