Deep learning with knowledge graphs using graph neural networks
Knowledge Graphs (KGs) are reshaping the paradigm of representing, organising, and utilising information about the world. They provide rich semantic information, and have emerged as a driving power of Artificial Intelligence (AI). Primarily, there are two types of important research directions for K...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1817931058101878784 |
|---|---|
| author | Liu, S |
| author2 | Horrocks, I |
| author_facet | Horrocks, I Liu, S |
| author_sort | Liu, S |
| collection | OXFORD |
| description | Knowledge Graphs (KGs) are reshaping the paradigm of representing, organising, and utilising information about the world. They provide rich semantic information, and have emerged as a driving power of Artificial Intelligence (AI). Primarily, there are two types of important research directions for KGs: one focuses on constructing and improving the quality of KGs, and the other delves into the wide range of applications of KGs. Recent years have also witnessed the advancement of Graph Neural Networks (GNNs), which are a class of deep learning techniques applicable to the graph domain and have demonstrated promising performance in many tasks. While there have been research attempts of applying GNNs to the KG-related tasks, there still remain several open challenges with the models’ function design, scalability issues, transductive nature of being limited to predicting for entities observed during training, and the quality of the benchmarks. In this thesis, towards deep learning with KGs using GNNs, we consider the tasks of inductive KG completion and inductive knowledge-enhanced recommendation in the context of the two directions, and propose novel GNN-based approaches to address the challenges. Our extensive empirical evaluation shows that our approaches outperform the state-of-the-art approaches on a collection of baselines, and can achieve efficient training and testing in practice. We also take a further step into the KG completion problem by revisiting the benchmarks in the transductive settings. In particular, we propose a new approach to generate benchmarks that can help empirically assess models’ ability to capture inference patterns. Our findings highlight the gaps between theoretical and empirical results concerning such inference ability. |
| first_indexed | 2024-12-09T03:15:59Z |
| format | Thesis |
| id | oxford-uuid:3a7e1f55-b03b-4ccf-9474-e8d035869c7f |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-12-09T03:15:59Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:3a7e1f55-b03b-4ccf-9474-e8d035869c7f2024-10-25T09:46:01ZDeep learning with knowledge graphs using graph neural networksThesishttp://purl.org/coar/resource_type/c_db06uuid:3a7e1f55-b03b-4ccf-9474-e8d035869c7fNeural networks (Computer science)EnglishHyrax Deposit2024Liu, SHorrocks, ICuenca Grau, BKostylev, EKnowledge Graphs (KGs) are reshaping the paradigm of representing, organising, and utilising information about the world. They provide rich semantic information, and have emerged as a driving power of Artificial Intelligence (AI). Primarily, there are two types of important research directions for KGs: one focuses on constructing and improving the quality of KGs, and the other delves into the wide range of applications of KGs. Recent years have also witnessed the advancement of Graph Neural Networks (GNNs), which are a class of deep learning techniques applicable to the graph domain and have demonstrated promising performance in many tasks. While there have been research attempts of applying GNNs to the KG-related tasks, there still remain several open challenges with the models’ function design, scalability issues, transductive nature of being limited to predicting for entities observed during training, and the quality of the benchmarks. In this thesis, towards deep learning with KGs using GNNs, we consider the tasks of inductive KG completion and inductive knowledge-enhanced recommendation in the context of the two directions, and propose novel GNN-based approaches to address the challenges. Our extensive empirical evaluation shows that our approaches outperform the state-of-the-art approaches on a collection of baselines, and can achieve efficient training and testing in practice. We also take a further step into the KG completion problem by revisiting the benchmarks in the transductive settings. In particular, we propose a new approach to generate benchmarks that can help empirically assess models’ ability to capture inference patterns. Our findings highlight the gaps between theoretical and empirical results concerning such inference ability. |
| spellingShingle | Neural networks (Computer science) Liu, S Deep learning with knowledge graphs using graph neural networks |
| title | Deep learning with knowledge graphs using graph neural networks |
| title_full | Deep learning with knowledge graphs using graph neural networks |
| title_fullStr | Deep learning with knowledge graphs using graph neural networks |
| title_full_unstemmed | Deep learning with knowledge graphs using graph neural networks |
| title_short | Deep learning with knowledge graphs using graph neural networks |
| title_sort | deep learning with knowledge graphs using graph neural networks |
| topic | Neural networks (Computer science) |
| work_keys_str_mv | AT lius deeplearningwithknowledgegraphsusinggraphneuralnetworks |