This repository includes a toy signal-and-noise HMM dataset and a basic implementation of the Tree-regularized GRU model in NumPy Autograd. Please find the ArXiv copy here: https://arxiv.org/abs/1711.06178.

For more on NumPy Autograd, see https://github.com/HIPS/autograd.

Create a new conda environment and activate it.

conda create -n interpret python=2
source activate interpret

Install the necessary libraries.

pip install -r requirements.txt

First, generate the toy dataset. This will create a directory ./data and populate it with Pickle files.

python dataset.py

Then, you can train the tree-regularized GRU. This will create a directory ./trained_models and dump the trained model and a PDF of the final decision tree. You can set the regularization strength as a command line argument. apl stands for average path length.

python train.py --strength 1000.0

We can apply the trained model on a held-out test set. This will print out AUC statistics.

python test.py

Please reach out with any concerns or potential bugs.