This repository contains code used to produce results showcased in the paper:
Graph Learning meets regularized factor models,
Florent Bouchard, Arnaud Breloy, Alexandre Hippert-Ferrer, Ammar Mian, Titouan Vayer
Article available at: https://arxiv.org/pdf/2210.11950.pdf
.
├── conda/
│ └── environment.yml
├── data/
│ ├── gps_up.npy
│ └── animal.mat
├── src/
│ ├── utils.py
│ ├── elliptical_estimation.py
│ ├── elliptical_graph_model.py
│ ├── estimators.py
│ ├── manifolds.py
│ ├── models.py
│ ├── sparse_penalties.py
│ └── visualization.py
├── tex/
│ ├── gnss_results.tex
│ ├── VIEW.GNSS_map.eps
│ └── VIEW.GNSS_map-eps-converted-to.pdf
├── ROC_curves.py
├── benchmark_animal.py
├── benchmark_gps.py
├── benchmark_concepts.py
├── graphical_models.py
├── install.sh
├── LICENCE
└── README.md
Most of the development is done as functions and classes available in the ./src/ directory. Simulations presented in the paper can be reproduced using the scripts ROC_curves.py, benchmark_animal.py, benchmark_gps.py, gnss_results.tex, benchmark_concepts.py. Data is stored in the ./data/ directory.
The simulationw were developped on linux system with python 3.10 with various libraries. In order to help with reproductibility matters, a conda environment was used. To obtain the same running environment, the best way is to install the same environment.
It is then necessary to have either miniconda or Anaconda available:
- Miniconda: https://docs.conda.io/en/latest/miniconda.html
- Anaconda: https://www.anaconda.com/products/distribution
-
Installing conda environment
To install the environment, a bash script install.sh is given in the root of this directory. Simply run:
./install.sh -
Activating conda environment and running a simulation
To activate environment, run:
conda activate graphfactormodelsThen running a simulation is easily done using python. For example, run:
python simulations/graphical_models.pyto generate and visualize the considered synthetic graph structures in the paper.
The following correspondance table shows the files used to produce all figures presented in the paper:
| File | Produces figure |
|---|---|
graphical_model.py |
(1) in Suppl. material |
ROC_curves.py |
(2)--(7) in Suppl. material |
benchmark_animal.py |
(1) in main paper |
benchmark_gps.py and /tex/gnss_results.tex |
(2) in main paper |
benchmark_concepts.py |
(8), (9) in Suppl. material |
Compute and visualize receiver operating caracteristics (ROC) curves. Three ROC curves are computed to perform the following:
- Sensitivity of the proposed approaches to the regularization parameter;
- Sensitivity of the proposed approaches to the rank of the factor model;
- Comparison of the proposed approaches to state-of-the art methods.
| Parameter | Type | Possible values | Description |
|---|---|---|---|
| method | str |
'GGM', 'EGM', 'GGFM', 'EGFM', 'all' (default: 'GGM') | Method(s) to use. 'all' performs a comparison of SOTA methods, which can only be used with roc='compare'. Other values are to used with roc='lambda' or 'roc=rank'. |
| graph | str |
'BA', 'ER', 'WS', 'RG' (default: 'BA') | Graph structure to be considered. |
| roc | str |
'lambda', 'rank', 'compare' (default: 'lambda') | Type of ROC curve to be computed. 'compare' can only be used with a value of method='all'. 'lambda' and 'rank' modes are to be used with method='GGM', 'EGM', 'GGFM' or 'EGFM'. |
| lambda_val | float |
Real positive value (default: 0.05) | Regularization parameter. |
| rank | int |
Positive integer (default: 20) | Rank of the factor model. |
| n | int |
Positive integer (default: 105) | Number of samples. |
| multi | bool |
True or False (default: True) |
Whether to use multi-threading or not. |
| save | bool |
True or False (default: False) |
Save plot in pdf format in a /results folder. |
ROC curve in pdf format.
python ROC_curves.py --method <method> --graph <graph> --roc <roc_type> --lambda_val <lambda_value>
--rank <rank> -n <number_of_samples> --multi <bool> --save <bool>
python ROC_curves --method EGM --graph ER -n 200 --roc rank
python ROC_curves --method EGM --graph ER --roc lambda
python ROC_curves --method all --graph ER -n 200 --roc compare
Performs a comparison of the proposed approach to state-of-the art methods on the animal data set. Graphs matrices (precision) are estimated and displayed using the pyvis library.
None.
Estimated graphs in html format. These graphs will automatically open in a browser (note that it has only been tested with Firefox and Chrome browsers). If the graph does not display, please refresh the page or try with another browser.
python benchmark_animal.py
Performs a comparison of the proposed approach to state-of-the art methods on the GNSS data set. Graphs matrices (precision) are estimated and displayed using the pyvis library.
None.
- Estimated graphs in
htmlformat. These graphs will automatically open in a browser (tested with Firefox and Chrome browsers). If the graph does not display, please refresh the page or try with another browser. - Graph edges in
datformat.
python benchmark_gps.py
Display estimated graphs of the GNSS data set on the Piton de la Fournaise map (copyright WEBOBS/IPGP). Takes the dat files produced by benchmark_gps.py as inputs.
None.
A 5-pages pdf file containing the graphs estimated with StudentGL, GGM, EGM, GGFM, EGFM (in this exact order).
pdflatex gnss_results.tex
Performs a comparison of the proposed approach to state-of-the art methods on the concepts data set. Graphs matrices (precision) are estimated and displayed using the pyvis library.
Note: authors are not authorized to make the concepts data set public. This data set can be shared on demand.
None.
Will throw an error because the concepts data is not loaded (see Description).
python concepts.py
For any issue, please contact: [mail a remplir](mail a remplir)
- Florent Bouchard
- Arnaud Breloy
- Alexandre Hippert-Ferrer
- Ammar Mian
- Titouan Vayer