ExKMC

This repository is the official implementation of ExKMC: Expanding Explainable k-Means Clustering.

We study algorithms for k-means clustering, focusing on a trade-off between explainability and accuracy. We partition a dataset into k clusters via a small decision tree. This enables us to explain each cluster assignment by a short sequence of single-feature thresholds. While larger trees produce more accurate clusterings, they also require more complex explanations. To allow flexibility, we develop a new explainable k-means clustering algorithm, ExKMC, that takes an additional parameter k' ≥ k and outputs a decision tree with k' leaves. We use a new surrogate cost to efficiently expand the tree and to label the leaves with one of k clusters. We prove that as k' increases, the surrogate cost is non-increasing, and hence, we trade explainability for accuracy.

Installation

The package is on PyPI. Simply run:

pip install ExKMC

Usage

from ExKMC.Tree import Tree
from sklearn.datasets import make_blobs

# Create dataset
n = 100
d = 10
k = 3
X, _ = make_blobs(n, d, k, cluster_std=3.0)

# Initialize tree with up to 6 leaves, predicting 3 clusters
tree = Tree(k=k, max_leaves=2*k) 

# Construct the tree, and return cluster labels
prediction = tree.fit_predict(X)

# Tree plot saved to filename
tree.plot('filename')

Notebooks

Usage examples:

• Expand tree for Gaussians
• Expand tree for text data

Citation

If you use ExKMC in your research we would appreciate a citation to the appropriate paper(s):

• For IMM base tree you can read our ICML 2020 paper.

  @article{dasgupta2020explainable,
    title={Explainable $k$-Means and $k$-Medians Clustering},
    author={Dasgupta, Sanjoy and Frost, Nave and Moshkovitz, Michal and Rashtchian, Cyrus},
    journal={arXiv preprint arXiv:2002.12538},
    year={2020}
  }
  

• For ExKMC expansion you can read our paper.

  @article{frost2020exkmc,
    title={ExKMC: Expanding Explainable $k$-Means Clustering},
    author={Frost, Nave and Moshkovitz, Michal and Rashtchian, Cyrus},
    journal={arXiv preprint arXiv:2006.02399},
    year={2020}
  }
  

Contact

• Nave Frost
• Michal Moshkovitz
• Cyrus Rashtchian