Modular framework for semi-supervised classification across heterogeneous data (text, vision, tabular, audio, graph).

Main repository · Documentation · Paper (arXiv) · Paper (HAL)

ModSSC targets research workflows where you need a single, reproducible pipeline to compare semi-supervised methods fairly.

Core ideas:

• Composable bricks: dataset loaders, sampling, preprocessing, graph construction, SSL methods, evaluation.
• Inductive and transductive settings share the same mental model and CLI surface.
• Reproducibility by design: declarative YAML experiments, deterministic seeds, fingerprinted caches, audit-friendly run folders.
• Extensibility: registries let you add datasets, samplers, preprocess steps, graph specs, and methods without rewriting the runner.

Current status: 0.0.1 (expect API changes until 1.0).

The project lives in the main repo:

• https://github.com/ModSSC/ModSSC

Quick start from source:

git clone https://github.com/ModSSC/ModSSC.git
cd ModSSC
python -m pip install -e ".[dev]"

Explore the CLI:

modssc --help
modssc-datasets list
modssc-preprocess steps list
modssc-graph specs list
modssc-inductive methods list
modssc-transductive methods list

Run end-to-end examples:

python examples/00_inductive_toy_self_training.py
python examples/01_transductive_toy_label_propagation.py

The bench/ runner is part of the GitHub repository (not intended as a packaged dependency).

python -m bench.main --config bench/configs/experiments/toy_inductive.yaml
python -m bench.main --config bench/configs/experiments/toy_transductive.yaml

Open issues and feature requests in the main repository:

• https://github.com/ModSSC/ModSSC/issues

This .github repository stores organization-wide community health files (issue templates, PR templates, policies).

If you use ModSSC in research, cite the project via CITATION.cff (and the paper linked above).

MIT.