Joint Probability Trees

Joint Probability Trees (JPTs) are a formalism for learning and reasoning about joint probability distributions that is tractable for practical applications. JPTs support both symbolic and subsymbolic variables in a single hybrid model without relying on prior knowledge about variable dependencies or families of distributions.

JPT representations build on tree structures that partition the probability space into relevant subregions elicited from training data, rather than postulating a rigid dependency model prior to learning. Learning and inference scale linearly, and the tree structure enables white-box reasoning about any posterior probability P(Q|E), providing interpretable explanations for every inference result.

Installation

Install the core package from PyPI:

pip install pyjpt

Optional dependencies

Depending on your use case, install one or more optional dependency groups:

pip install pyjpt[matplotlib]   # matplotlib and graphviz plotting
pip install pyjpt[plotly]       # interactive plotly plotting
pip install pyjpt[seq]          # sequential/temporal models
pip install pyjpt[mlflow]       # MLflow experiment tracking

Multiple groups can be combined:

pip install pyjpt[matplotlib,mlflow]

Development setup

Clone the repository and install in editable mode with all development dependencies:

git clone https://github.com/joint-probability-trees/jpt-dev
cd jpt-dev
pip install -e ".[dev]"

By default, Cython extensions are compiled during installation. For development, you can skip pre-compilation and let pyximport handle on-the-fly compilation at runtime instead:

JPT_NO_CYTHON=1 pip install -e ".[dev]"

With on-the-fly compilation, changes to .pyx files are picked up automatically on the next import without requiring a rebuild.

Running tests

cd test
python -m unittest discover

Building distributions

python -m build            # sdist + wheel
python -m build --sdist    # source distribution only
python -m build --wheel    # wheel only

Documentation

The documentation is hosted on Read the Docs.

Building the documentation locally

Install the documentation dependencies:

pip install -r doc/requirements.txt

Then build the HTML documentation from the doc/ directory:

cd doc
make html

The output is written to doc/build/html/. Open doc/build/html/index.html in a browser to view it. The html target automatically cleans previous build artifacts before rebuilding.

Citation

If you use JPTs in your research, please cite:

@inproceedings{nyga23jpts,
    title     = {{Joint Probability Trees}},
    author    = {Daniel Nyga and Mareike Picklum and
                 Tom Schierenbeck and Michael Beetz},
    year      = {2023},
    booktitle = {arxiv.org},
    url       = {http://arxiv.org/abs/2302.07167}
}