What Drives Schema Capture in Link Prediction?

Overview & Synopsis

Due to their semi-automatic construction process, Knowledge Graphs (KGs) inherently suffer from incompleteness issues, addressed by Link Prediction approaches. The most promising ones are based on Knowledge Graph Embedding models (KGEMs) and are traditionally evaluated on their abilities to predict true triples. However, recent works explored other evaluation aspects of KGEMs such as the Sem@k metric that assesses the ability of a model to predict coherent triples w.r.t. the KG schema, in particular the domain and range of relations. However, previous experiments were limited to datasets with complete schemas, i.e. containing only typed entities and relations with both a domain and a range. In this work, we investigate the schema capture of a range of KGEMs on realistic datasets with incomplete schemas, identifying factors that influence a model to learn the schema. Our results show that datasets with incomplete schema bring additional challenges to KGEM schema capture; performance is linked to the number of train triples per relation, as well as to specific KGEM architectures.