MoAT: Mixtures of All Trees

This repository contains the implementation of the MoAT model presented in our AISTATS23 paper titled Mixtures of All Trees. MoAT is a mixture over all possible $n^{(n-2)}$ tree-shaped graphical models over $n$ variables.

Usage

• Install dependencies from requirements.txt.
• Run ./learn.sh with the appropriate dataset name and hyperparameters of your choosing.

Notes

• The collection of datasets used in our density estimation experiments is provided in the datasets folder. You may import your own datasets too. The current implementation requires that the datasets be binarized.
• The source code for our sampling experiments is availble in sample.py, and the experiments can be conveniently invoked with ./sample.sh. Further, the methods in the MoAT class in models.py can be leveraged to obtain the samples directly too.