Beyond Mapping : Domain-Invariant Representations via Spectral Embedding of Optimal Transport Plans

This repository contains the official implementation of the paper "Beyond Mapping : Domain-Invariant Representations via Spectral Embedding of Optimal Transport Plans", authored by Abdel Djalil Sad Saoud, Fred Maurice Ngolè Mboula, and Hanane Slimani. Accepted at the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) [paper].

Table of Contents

• Overview
• Installation
• Run
• Citation
• References

Overview

This paper presents a novel Optimal Transport-based domain adaptation method that interprets optimal transport plans as adjacency matrices capturing cross-domain connectivity, then derives domain-invariant and discriminative representations of samples via spectral embedding [1].

Key contributions:

1. Novel OT-based framework: We propose a domain adaptation approach that leverages the cross-domain connectivity captured by the transport plans to compute domain-invariant and discriminative representation of samples, rather than estimating a mapping from one domain to another in the samples space.

2. Multi-source extension: We extend our framework to multi-source domain adaptation scenarios, inspired from [2].

3. Comprehensive evaluation: We evaluate our method on acoustic adaptation benchmarks and demonstrate industrial relevance through a cable defect diagnosis use case based on Time Domain Reflectometry.

Installation

Development Environment:

• Python 3.10.0
• pip 25.3

Install all required packages:

pip install -r requirements.txt

Run

To run SeOT

Step 1. Prepare Data

Download datasets to the data/ directory.

Step 2. Run Experiments

For MSD/MGR benchmarks:

# On MSD
python main_seot_msda.py --benchmark MSD --algorithm SeOT --reg_e 1e-4 --reg_e_bar 1e-2 --n_component 10 --epochs 100 --lr 0.001 --batch-size 128

# On MGR
python main_seot_msda.py --benchmark MGR --algorithm SeOT --reg_e 1e-4 --reg_e_bar 1e-2 --n_component 10 --epochs 100 --lr 0.001 --batch-size 128

Citation

If you use this code or find our work useful for you research, please consider citing our paper:

@misc{saoud2026mappingdomaininvariantrepresentations,
      title={Beyond Mapping : Domain-Invariant Representations via Spectral Embedding of Optimal Transport Plans}, 
      author={Abdel Djalil Sad Saoud and Fred Maurice Ngolè Mboula and Hanane Slimani},
      year={2026},
      eprint={2601.13350},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2601.13350}, 
}

References

[1] Scikit-learn developers. sklearn.manifold.SpectralEmbedding. In scikit-learn Documentation. Retrieved from https://scikit-learn.org/stable/modules/generated/sklearn.manifold.SpectralEmbedding.html.

[2] Montesuma, E. F. & Mboula, F. M. N. (2021). Wasserstein Barycenter Transport for Multi-Source Domain Adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).