The codes have been tested in Mathworks Matlab R2021a. We used Matlab (The MathWorks, Inc.) R2021a. To set up the environment, download the following toolboxes:
After downloading the toolboxes, modify the paths in the Matlab scripts accordingly.
The code in the contour-plots folder uses the GPML toolbox for MCMC estimation of marginal likelihood surface. The scripts run_lml.sh and run_lp.sh were used for submitting the jobs for log marginal likelihood estimation and log predictive density estimation. These scripts produce the results for the ionosphere data set (leftmost panels in Fig. 2 in the paper).
Surface contour plot jobs for the other data sets can be submitted by changing the data set name in the bash scripts.
The code in the classification folder uses the GPStuff toolbox for MCMC log predictive density estimation in classification. The script run_mcmc_classification.sh was used for submitting the jobs.
To run these scripts, we use the same splits as in Python. To ensure the splits to be the same, we export the pre-split data and write it on disk. These data-splits can be created by running
python generate_matlab_data.pyin the python/experiments/classification folder (found in the root of this repository). Alternatively, the generated splits can be downloaded as mat files from Google Drive as a zip file. Unzip the file inside the classification folder.
The code in the student-t folder uses the GPStuff toolbox for LA, EP and MCMC log predictive density estimation in Student-t regression. The script run_mcmc_student_t.sh was used for submitting the jobs.