JDCOOT : A Joint Distribution Co-Optimal Transport Approach for Adapting Models to Heterogeneous Domains

Pierre Navaro (CNRS), Valérie Garès (INRIA) and Chloé Friguet (UBS) with contribution from Lucas Offroy and Marion Jeamart (Internships)

This work addresses a fundamental challenge in modern statistical learning: adapting models to heterogeneous domains, where source and target data are characterised by different feature spaces and underlying distributions. We introduce Joint Distribution Co-Optimal Transport (JDCOOT), a domain adaptation algorithm that leverages optimal transport to align the joint feature-label distributions of distinct domains, enabling effective knowledge transfer across heterogeneous domains.

Installation

Use pixi to run the code

curl -fsSL https://pixi.sh/install.sh | bash
echo 'eval "$(pixi completion --shell bash)"' >> ~/.bashrc
source ~/.bashrc

All run scripts are in the examples directory:

git clone https://github.com/otrecoding/jdcoot/
cd jdcoot
pixi install
pixi run python examples/discrete_partial_jdcoot.py

Numerical experiments

For all expermiments we use three methods reference, coot and jdcoot on two datasets (train and test). The "pure" performance is the accuracy on the train dataset and the "test" performance is the prediction accuracy using the same model on the test dataset.

Input :

learning data, 2 dataframes (source and target) with the following format :

$$ X^S = (X 1^S . . . X d_S^S) Z \qquad \mbox{and} \qquad X^T = (X 1^T \ldots X d_T^T) Z $$

• $X$ are the observed covariate,
• $Z$ the discrete objective variable for classification analysis

test data, 2 dataframes (test_source and test_target)

supervision :

• unsupervised : none of the observations of train target are labelled and all source observations are labelled
• semi-supervised : prop_target of the observations of train target are labelled and all train source observations are labelled
• partial : prop_target of the observations of train target are labelled and prop_source of the observations of train source are labelled