Dependencies

Python3.7.3 and some python3 libraries:

• numpy (v1.16.2 used)
• sklearn (v0.22.2.post1 used)
• scipy (v1.2.1 used)
• lightgbm (v2.3.1 used)
• matplotlib (v3.0.3 used)

Content

folder 'datasets'

• contain the 16 datasets used in the experiments

folder 'experiment-toy-LGBM-GBRFF2':

• contain the source code to reproduce the experiment on toy datasets where LGBM and our proposed method GBRFF2 are compared
• the experiment is launched with 'python main.py seed' which apply the algorithms on the generated datasets with a fixed seed, and store the results in a file. We executed this script with seeds from 1 to 20.
• the file 'plotBoundaryAccuracy.py' load the result files, and produce the figure used in the paper (some modifications were made to produce the exact same figure in the paper)

folder 'experiment-progression-PBRFF-GBRFF1-GBRFF2':

• contain the source code to reproduce the experiments showing the progression between PBRFF, GBRFF0.5, GBRFF1, GBRFF1.5 and GBRFF2
• the experiment is launched with 'python main.py seed' which apply the algorithms on the 15 first datasets from the dataset folder with a fixed seed, and store the results in a file. We executed this script with seeds from 1 to 20.
• the file 'plot.py' load the result files, and produce the figures used in the paper

folder 'experiment-latex-array-results':

• contain the source code to reproduce the experiments producing a latex array comparing BMKR, GFC, PBRFF, GBRFF1, LGBM and GBRFF2
• the experiment is launched with 'python main.py seed' which apply the algorithms on the 15 first datasets from the dataset folder with a fixed seed, and store the results in a file. We executed this script with seeds from 1 to 20.
• To launch the experiment on the last dataset "bankmarketing" the file "datasets.py" needs to be modified (at the very bottom of the file)
• the file 'latex.py' load the result files, and produce a .tex file containg a tabular used in the paper

folder 'experiment-computation-time':

• contain the source code to reproduce the experiments comparing the computation time of the methods on an artificially generated dataset having an increasing number of examples
• the experiment is launched with 'python main.py seed' which apply the algorithms on the generated datasets until the time limit is reached with a fixed seed, and store the results in a file. We executed this script with seeds from 1 to 20.
• the file 'plot.py' load the result files, and produce the figure used in the paper

folder 'experiment-toy-loss-placing-landmark-GBRFF1':

• contain the source code to reproduce the toy experiment where we show on the two moons datasets that learning a single scalar is equivalent to learn a full landmark vector to minimize our loss function
• the file 'main.py' launch the experiment and produce the figure used in the paper

folder 'experiment-bars-GBRFF1':

• contain the source code to reproduce the experiment comparing GBRFF1 with different total number of random features T*K in the form of a barchart
• the experiment is launched with 'python main.py seed' which apply the algorithms on the 15 first datasets from the dataset folder with a fixed seed, and store the results in a file. We executed this script with seeds from 1 to 20.
• the file 'bars.py' load the result files, and produce the figure used in the paper