mdl-fca

Learning "good" concept DAGs from binary data by minimum description length — a probabilistic/information-theoretic reworking of Formal Concept Analysis.

Plain FCA maps a binary object×attribute context to a concept lattice, but it is not a probabilistic model: it creates a concept for every random coincidence, so the lattice explodes. This project builds instead the small DAG of concepts that best compresses the data: a concept exists only if it pays for its own description. Concepts can have multiple parents (unlike hierarchical clustering's trees) and are organized hierarchically (unlike flat topic models).

Status: design phase complete, prototype under construction.

• Read CLAUDE.md for orientation and design commitments.
• Read docs/ in numeric order for the full design:
  1. background and goal
  2. model and codelength (the core)
  3. algorithm (greedy pair-merge constructor + search framework)
  4. batch vs online
  5. prototype spec (current build target)
  6. roadmap

Quick start (once the prototype exists)

pip install -e ".[test]"
pytest
python examples/run_planted.py