End-to-End Training of Unsupervised Trees: KAURI and DOUGLAS

Louis Ohl, Pierre-Alexandre Mattei, Arnaud Droit, Frederic Precioso

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Decision trees, Clustering, Unsupervised Learning, end-to-end learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a framework for training end-to-end trees for clustering

Abstract:

Trees are convenient models for obtaining explainable predictions on relatively
small datasets. While many proposals exist for end-to-end construction of such
trees in supervised learning, learning a tree end-to-end for clustering without la-
bels remains an open challenge. As most works focus on interpreting with trees
the result of another clustering algorithm, we present here two novel end-to-end
trained unsupervised trees for clustering, respectively KAURI for datasets with a
large number of features using binary decision trees, and DOUGLAS for datasets
with a large number of samples using k-ary differentiable trees. Both methods are
composed of a learnable tree structure in which parameters are optimised accord-
ing to a generalised mutual information (GEMINI) and present results on par with
other existing methods while maintaining interpretability. We compare these two
models on multiple datasets with the most recent unsupervised trees and provide
guidelines for choosing the most suitable model.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5658