Graph signal processing techniques for machine learning optimization
Graph Signal Processing (GSP) offers a flexible framework for extending classical signal processing techniques, such as filtering and sampling, to irregular domains like social networks. In this context, the Q-GFT, a generalization of the Graph Fourier Transform (GFT) that incorporates a non-trivial...
| Autor: | |
|---|---|
| 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:2117/418112 |
| Acceso en línea: | https://hdl.handle.net/2117/418112 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine learning Neural networks (Computer science) Signal processing graph graph signal processing machine learning neural networks Aprenentatge automàtic Xarxes neuronals (Informàtica) Tractament del senyal Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal |
| Sumario: | Graph Signal Processing (GSP) offers a flexible framework for extending classical signal processing techniques, such as filtering and sampling, to irregular domains like social networks. In this context, the Q-GFT, a generalization of the Graph Fourier Transform (GFT) that incorporates a non-trivial inner product, was introduced to provide a more flexible analysis of graph signals. The Q-GFT brings additional flexibility through its variation operator, graph partitioning strategies, and modification of spectral properties. In this thesis, we explore the integration of the Q-GFT with graph neural networks (GNNs) for node classification tasks. We investigate the impact of Q-GFT and evaluate their effect on well-known GNN architectures such as Graph Convolutional Networks (GCNs) and GraphSAGE. Results provide valuable insights into the interaction between graph partitioning strategies and neural networks, showing how energy concentration patterns from different Q-based Graph Fourier Transforms (Q-GFT) correlate with model performance. While most models did not outperform the baseline, using ground-truth labels for the Fully Balanced max-cut did, suggesting potential for new partitioning methods to advance graph machine learning. |
|---|