Graph Learning
Supporting material for our paper accepted for presentation at EUSIPCO 2023, and available on HAL and arXiv.
Authors
- Benjamin Girault, ENSAI, FRANCE (Corresponding Author)
- Eduardo Pavez, USC, USA
- Antonio Ortega, USC, USA
Experiment Analysis
Supporting material comparing Pavez et al. 2019 and our proposed approach can be found below.
Metric Comparison
These figures include all 50 vertex samplings. The first two figures below are larger, and not included in the page to avoid web client overloading.
- Relation between edge weight vertex distance, including theoretical upper bounds
- Edge weight CDF
- Relation between vertex importances and closest vertex distance
Vertex importance vs closest vertex distance
- Graph Power Spectrum Densities (gPSD)
Graph power spectrum densities
Learnt Graphs
The graphs below are all from the same sampling of the space (one set of vertex locations), with different ranges for the variogram, and for both Pavez et al. 2019, and our proposed approach. Edges are trimmed to those of weights greater than 1e-6 before being displayed (justified above using the edge weight CDF). Edges are then binned, and each bin is associated to a shade of gray.