MATLAB implementation of an algorithm for source localization on graphs via l1 regularization and spectral graph theory. As presented in our paper:
R. Pena, X. Bresson, and P. Vandergheynst. "Source Localization on Graphs via l1 Regularization and Spectral Graph Theory". 12th IEEE Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), 2016.
The code is released under the terms of the MIT license. Please cite the paper above if you use it.
Make sure to download and install the following toolboxes:
- Clone this repository.
git clone https://github.com/rodrigo-pena/src-localization-graphs
cd src-localization-graphs-
Install the dependencies mentioned on the previous section.
-
In MATLAB, change the current directory to the src-localization-graphs folder and run src_loc_start.m to automatically add all the subfolders to your path.
-
Run demo_alternate_optimization.m to see if everything works fine.
If you want to work with John Snow's GIS, or the ETEX data (as used in the paper), you will need to clone another repository, load-data, and follow the installation instructions in the README therein.
Contains the main functions of the toolbox. Use function alt_opt.m to simultaneously learn the source locations and the diffusion kernel. If the source locations are known, use learn_param_kernel.m to learn the diffusion kernel. If the diffusion kernel is known, use learn_sparse_signal.m to learn the source locations.
Contains a third party function to plot error bars over a surface embedded in 3-dimensional space. Used only in testbench.m.
Contains demonstrations of what can be done with the toolbox. The names should be self-explanatory.
Contains the scripts for reproducing the experiments in paper.
Contains the numerical solvers used in the alternate optimization problem: FISTA, and Newton's method.
Contains useful functions for the toolbox, such as error measures, parameter estimation and initialization strategies, etc.
To use the source localization algorithms on your data, you'll need:
- A graph stored in GSPBox-compatible format.
- A source indicator vector, with non-zero elements on the indices representing the source nodes, and zeros everywhere else.
- An observation vectorn whose entries encode the signal value at each node.
- A parametric kernel representing the diffusion process.
See any of demos for examples on how to call the methods. Please get in touch if you are unsure about how to adapt the code to your settings.
R. Pena, X. Bresson, and P. Vandergheynst. "Source Localization on Graphs via l1 Regularization and Spectral Graph Theory". 12th IEEE Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), 2016.