!!!DISCLAIMER!!!

A better and more recent Julia implementation exists now at : AugmentedGaussianProcesses.jl , this repository only contains an older version of the algorithm as well the experiments presented in our paper, it is relying on an outdated version of Julia

README

Objective

• This repository contains the updated source code for the Bayesian Nonlinear Support Vector Machine (BSVM) both in its stochastic (and with inducing points) and its batch version
• It relates to the paper published at the conference track of ECML 17' "Bayesian Nonlinear Support Vector Machines for Big Data" by Florian Wenzel, Théo Galy-Fajou, Matthäus Deutsch and Marius Kloft. Paper is available at https://arxiv.org/abs/1707.05532

How do I install the package?

• First clone this repository (git clone https://github.com/theogf/BayesianSVM.git)

• If you simply want to try out the package you need to install the Julia dependencies :

  • Distributions
  • PyPlot
  • StatsBase
  • GaussianMixtures
  • Clustering
  • ScikitLearn

  Note: to install new packages use the Pkg.add("ModuleName") function in Julia

• If you want to try the competitors as well you will need to install these Julia and Python dependencies (as well as Python ofc):

  • (Julia)PyCall
  • (Python)ScikitLearn
  • (Python)Tensorflow
  • (Python)GPflow

  Note: to use Tensorflow and GPflow, they must me included in the search path of PyCall, to do this use : unshift!(PyVector(pyimport("sys")["path"]), "path_to_add") and call Pkg.build("PyCall"), also note that they are much more complicate to install

• Both tests and source files are written in Julia (v0.5), one first needs to julia to run those, however a Python or Matlab user should be able to read easily through the code as the syntax is quite similar

• Some light datasets are included (especially the Rätsch Benchmark dataset), the SUSY dataset can be found on UCI

How to run tests?

• Go to the "test" folder, open "run_test.jl", chose the dataset and change the parameters (more is explained in the file) and simply run the file. (for example change the type of BSVM (linear/nonlinear, sparse, use of stochasticity etc)
• If you want to also use the competitors, open "paper_experiments.jl", chose the dataset, chose the methods you want to test and adapt the parameters (more details in the file).
• For more custom usage of the BSVM method, look at the source code of src/BSVM.jl, where all the options are explained. More documentation will be there soon.

Who to contact

For any queries please contact theo.galyfajou at gmail.com