Exploiting Subgraphs and Attributes for Representation Learning on Knowledge Graphs

Knowledge graphs (KGs) are data structures that explicitly represent entities and the relations between them over some domain. They can be used to store information about people and their relatives or birth locations, about organizations and the countries where they are located, or about chemical compounds and their interactions with proteins in the human body. In this thesis, we investigate the problem of representation learning on knowledge graphs, which consists of learning vector representations of entities and relations that capture the information contained in the graph. These learned representations are useful for capturing patterns that occur in the graph but are not explicitly stated in it. Several methods for representation learning on KGs are based on predicting a link between a pair of entities in the graph. While efficient, this approach forgoes useful learning signals from other sources that exhibit patterns, such as subgraphs involving multiple entities, and attributes of entities. Examples of attributes are textual descriptions of people, or molecular structures of chemical compounds. The goal of this thesis is to explore such learning signals beyond pairwise interactions of entities. Our findings provide evidence that subgraphs and attributes are powerful signals from which we can learn representations in KGs. Not only do they yield improved representations, but they also broaden the range of tasks in which they can be applied. We hope that this serves as a motivation for learning from further sources of information that are already available in KGs, but whose potential is yet to be discovered.