Deep embeddings and Graph Neural Networks: can context improve domain-independent predictions?

Graph neural networks (GNNs) are deep learning architectures that apply graph convolutions through message-passing processes between nodes, represented as embeddings. GNNs have recently become popular because of how they obtain a contextual representation of each node that takes into account the inf...

Descripción completa

Detalles Bibliográficos
Autor: Sola Espinosa, Fernando Luis
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2022
País:España
Institución:Universidad de Sevilla (US)
Repositorio:idUS. Depósito de Investigación de la Universidad de Sevilla
OAI Identifier:oai:idus.us.es:11441/148087
Acceso en línea:https://hdl.handle.net/11441/148087
Access Level:acceso abierto
Palabra clave:Knowledge Graphs
Graph Neural Networks
Attributive embeddings
Deep graph embeddings
Machine Learning
Grafos de Conocimiento
Redes Neuronales de Grafos
Embeddings Atributivos
Embeddings profundos de grafos
Descripción
Sumario:Graph neural networks (GNNs) are deep learning architectures that apply graph convolutions through message-passing processes between nodes, represented as embeddings. GNNs have recently become popular because of how they obtain a contextual representation of each node that takes into account the information from surrounding nodes. Existing work has focused on the development of GNN architectures, using basic domain-specific information about the nodes to compute embeddings. In the context of knowledge graphs, much effort has been put towards developing deep learning techniques to obtain node embeddings that preserve information about relationships and structure without relying on domain-specific data. The potential of the application of graph neural networks with deep embeddings of knowledge graphs remains largely unexplored. In this project, we carry out a number of experiments to answer open research questions about how said embeddings perform when using a graph neural network. We test 7 different deep embeddings across several attribute prediction tasks in two attribute-rich datasets. We conclude that there is a significant performance improvement but it varies heavily depending on the task and deep embedding technique. Given the interest of the results obtained, we have submitted an article to the conference CIKM’22, and we have defined some tasks and future work in order to continue this study in the form of a PhD.