Towards a Data-Driven Graph Neural Network Model Selection

Graph Neural Networks (GNNs) have achieved great success in solving many machine learning tasks, and many different neural architectures have been proposed over the past few years. However, the lack of robust public benchmarks hinders the assessment of GNN architectures, which has made most research...

Descripción completa

Detalles Bibliográficos
Autor: Higueras Serrano, Raúl
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/411416
Acceso en línea:https://hdl.handle.net/2117/411416
Access Level:acceso abierto
Palabra clave:Neural networks (Computer science)
Machine learning
Graph theory
graph neural networks
neural networks
meta-learning
machine learning
deep learning
artificial intelligence
model selection
neural architecture search
graph theory
graph model
graph generator
data science
Xarxes neuronals (Informàtica)
Aprenentatge automàtic
Grafs, Teoria de
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descripción
Sumario:Graph Neural Networks (GNNs) have achieved great success in solving many machine learning tasks, and many different neural architectures have been proposed over the past few years. However, the lack of robust public benchmarks hinders the assessment of GNN architectures, which has made most research papers rely on the same 5-10 datasets. Our first contribution is Graphlaxy++, a novel generative graph learning model designed to emulate GNN performance and that can be used to generate artificial benchmarks with user-defined metrics. Leveraging Graphlaxy++, we generate a diverse synthetic graph dataset that serves as input to a predictive statistical model that can predict the performance of GNNs. This model is able to successfully predict the relative performance of 4 different architectures: GCN, GAT, GIN, and MLP. The predictive capacity of this model offers potential applications in speeding up neural architecture searches and optimizing graph modeling strategies.