Generalising Random Forest Parameter Optimisation to Include Stability and Cost

The code repository for the paper to appear in ECML-PKDD 2017:

Liu, C.H.B., Chamberlain, B.P., Little, D.A. and Cardoso, A., 2017. Generalising Random Forest Parameter Optimisation to Include Stability and Cost. arXiv preprint arXiv:1706.09865.

In BibTeX:

@article{liu2017generalising,
  title={Generalising Random Forest Parameter Optimisation to Include Stability and Cost},
  author={Liu, CH Bryan and Chamberlain, Benjamin Paul and Little, Duncan A and Cardoso, Angelo},
  journal={arXiv preprint arXiv:1706.09865},
  year={2017}
}

Setup

We provide two scripts to set up the project environment and load the public datasets.

OS utilities/ Anaconda

The scripts assumes the following utilities exists:

• gcc
• unzip
• anaconda

MacOSX users should run the following if the above is not available in their machine:

brew install gcc
brew install unzip
curl https://repo.continuum.io/archive/Anaconda2-4.3.1-MacOSX-x86_64.sh > ./local_resources/anaconda2_install.sh
bash ./local_resources/anaconda2_install.sh -b
rm -f ./local_resources/anaconda2_install.sh

Linux users should run the following:

sudo apt-get install gcc
sudo apt-get install unzip
curl https://repo.continuum.io/archive/Anaconda2-4.3.1-Linux-x86_64.sh > ./local_resources/anaconda2_install.sh
bash ./local_resources/anaconda2_install.sh -b
rm -f ./local_resources/anaconda2_install.sh

Project Environment Setup

The following script creates a separate environment gft_env in anaconda, and installs Python 2.7 and the necessary packages to run the experiment code.

./setup_environment.sh

The script requires ~500MB of network traffic and ~2G disk space.

Loading Public Datasets

The following scipt loads the public datasets from the internet. It assumes curl, unzip, and tar is installed in the machine already.

./load_data.sh

It loads five files into the local_resources directory:

• orange_small_train.data (The features for the Orange small dataset)
• orange_small_train_appetency.labels (The appetency labels for the Orange small dataset)
• orange_small_train_churn.labels (The churn labels for the Orange small dataset)
• orange_small_train_upselling.labels (The upselling labels for the Orange small dataset)
• criteo_train.txt (The features and labels for the Criteo dataset)

The script requires ~4.5G of network traffic and ~11-12G of disk space in addition to the setup script above.

Running the code

To run the Jupyter notebooks containing the project code:

source activate gft_env
jupyter notebook

There are four notebooks living in the src directory:

• BayesOpt.ipynb (For Bayesian optimisation experiments)
• Figure1.ipynb
• Figure2.ipynb
• Figure3.ipynb (For generating figures 1-3 in the paper)

Cleaning Up

After running the code, you can terminate the Jupyter notebook server, and leave the project environment by issuing the following command:

source deactivate

To remove the project environment, use the following command:

conda remove --name gft_env --all

Documentation

You can find the full derivation on MSPD being an empirical estimate (capturing the interaction between the variance of the random forest and the covariance of predictions between runs) in the doc directory.