(Attributed) Graph Partitioning with semi-relaxed (Fused) Gromov-Wasserstein

This repository implements simulation and clustering methods based on semi-relaxed (Fused) Gromov–Wasserstein (sr(F)GW) and k‑Fréchet means for attributed and non-attributed graphs.

Main functionalities

• Simulation of synthetic graphs (SBM-like mixing matrices) and node attributes (functional data and histograms).
• Distance computation: graph shortest-path distances, DTW for functional data, and Wasserstein for histograms.
• Clustering implementations: k‑Fréchet means, embedded k‑means, Semi-Relaxed GW (srGW) and Semi-Relaxed Fused GW (srFGW).
• Evaluation for automated experiments and metrics (internal/external).
• Example notebooks that reproduce experiments (located in top-level notebooks and methods_example.ipynb).

Repository layout

• src/: project Python package
  • src/simulation: graph and attribute simulators (graph.py, attributes.py, utils.py)
  • src/clustering: clustering algorithms and helpers (gw_clustering.py, kmeans.py, init_strategies.py, utils.py)
  • src/evaluation: evaluation utilities (utils.py, metrics.py, testing.py)
  • src/distances.py: distance computation helpers
  • src/shapes.py: SBM creation with constraints
• methods_example.ipynb: example workflows
• simulations.ipynb: simulations for (non)-attributed graph clustering evaluation
• figure_table_generation.ipynb: generation of figures used in the article

Requirements & install

• Python 3.8+ recommended.
• Install dependencies (in a virtual environment):

  python -m venv .venv
  source .venv/bin/activate
  pip install -r requirements.txt

Notable dependencies: numpy, scipy, pandas, scikit-learn, networkx, matplotlib, seaborn, POT (ot), dtaidistance, dtw.