Structural Alignment in Link Prediction

Abstract

While Knowledge Graphs (KGs) have become increasingly popular across various scientific disciplines for their ability to model and interlink huge quantities of data, essentially all real-world KGs are known to be incomplete. As such, with the growth of KG use has been a concurrent development of machine learning tools designed to predict missing information in KGs, which is referred to as the Link Prediction Task. The majority of state-of-the-art link predictors to date have followed an embedding-based paradigm. In this paradigm, it is assumed that the information content of a KG is best represented by the (individual) vector representations of its nodes and edges, and that therefore node and edge embeddings are particularly well-suited to performing link prediction. This thesis proposes an alternative perspective on the field's approach to link prediction and KG data modelling. Specifically, this work re-analyses KGs and state-of-the-art link predictors from a graph-structure-first perspective that models the information content of a KG in terms of whole triples, rather than individual nodes and edges. After building up a theoretical foundation for this structure-first approach from the state-of-the-art literature, it is evaluated in two contexts. The first evaluation asks if link predictors' outputs are aligned to aspects of KG structure. Results indicate that, not only are link predictors heavily influenced by structure, but that their patterns of hyperparameter preference, and their overall performance, can be explained and simulated in terms of the structure of the graph they were trained to learn. The second evaluation builds upon this observation and asks if graph structural features of triples in a KG are sufficient to enable link prediction. The results of this second round of experiments indicate that structure-based link prediction is not only possible, but highly effective compared to state-of-the-art approaches. Finally, it is has been found that, by representing the information content of a KG in terms of triple-level structure, cross-KG (including cross-domain) transfer learning becomes viable for the link prediction task. The thesis concludes that a structure-first perspective on KGs and link prediction is both viable and useful for understanding KG learning. This observation is used to create and propose the Structural Alignment Hypothesis, which postulates that link prediction can be understood and modelled as a structural task. All code and data used for this thesis, including the link prediction simulator (TWIG) and the structure-based link predictor (TWIG-I) are open-sourced to encourage further work in this area. Finally, this thesis was written bilingually, with the main document in English and an informal extended summary in Irish. An Irish-language translation dictionary of machine learning terms (the Foclóir Tráchtais) created for this work is open-sourced as well.