Differentiable Clustering with Perturbed Spanning Forests

TL;DR

This paper presents a novel differentiable clustering technique using stochastic perturbations of spanning forests, enabling end-to-end training and effective learning from partial data in noisy and complex geometries.

Contribution

It introduces a differentiable clustering method based on perturbed spanning forests, allowing integration into trainable pipelines and learning from partial clustering data.

Findings

Performs well on noisy and complex datasets

Effective in supervised and semi-supervised tasks

Enables end-to-end training with efficient gradients

Abstract

We introduce a differentiable clustering method based on stochastic perturbations of minimum-weight spanning forests. This allows us to include clustering in end-to-end trainable pipelines, with efficient gradients. We show that our method performs well even in difficult settings, such as data sets with high noise and challenging geometries. We also formulate an ad hoc loss to efficiently learn from partial clustering data using this operation. We demonstrate its performance on several data sets for supervised and semi-supervised tasks.