Wasserstein k-means++ for Cloud Regime Histogram Clustering

This repository contains the supporting code for the paper:

Staib, Matthew and Jegelka, Stefanie. Wasserstein k-means++ for Cloud Regime Histogram Clustering. In Proceedings of the Seventh International Workshop on Climate Informatics, 2017.

@inproceedings{staib2017wasserstein,
 author = {Staib, Matthew and Jegelka, Stefanie},
 title = {Wasserstein k-means++ for Cloud Regime Histogram Clustering},
 booktitle = {Proceedings of the Seventh International Workshop on Climate
Informatics: CI 2017},
 year = {2017}
}

Dependencies

• TFOCS
• Gurobi (for the Wasserstein gradient oracle)
• MOSEK (for an alternative Wasserstein gradient oracle)
• FastEMD (for an alternative Wasserstein gradient oracle)
• sinkhornTransport.m from Marco Cuturi (for an alternative Wasserstein gradient oracle)
• ICCSP D1 dataset

Getting started

1. First parse the ICCSP dataset by navigating to the directory with all the .hdf files, then running extract_cloud_histograms
2. Add the compute-optimal-transport directory to the path, as well as any other third party code (e.g. TFOCS)
3. Run cluster_histograms which will load the preprocessed ICCSP data and run various clustering algorithms. (be sure to modify the first few lines of cluster_histograms to load the .mat file from step 1, wherever it was stored)

At this point, the figures from the paper can be generated by running

• weather_state_plots
• prepare_globe_plots followed by make_globe_plots