flatironhealth/SUDO

Code to conduct the experiments outlined in the SUDO paper

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

SUDO

This repository contains the code required to conduct the experiments presented in the SUDO paper. Please note that the datasets to conduct such experiments are not provided here. If the datasets are publicly available, you can find links to them below.

Datasets

The experiments are conducted on several datasets:

1. Multi-Domain Sentiment (https://www.cs.jhu.edu/~mdredze/datasets/sentiment/index2.html)
2. Stanford DDI (https://ddi-dataset.github.io/index.html#access)
3. Camelyon17-WILDS (https://wilds.stanford.edu/get_started/)
4. Simulations

Conducting experiments

To conduct the SUDO experiments, you need to follow two steps:

• Step 1 - Download the data of interest (from above links)
• Step 2 - Extract features (or prediction probabilities) from the data
• Step 3 - Perform SUDO experiments

Step 2 - Feature extraction

To extract features from the datasets, please refer to the scripts entitled extract_XXX_features.py where XXX is a particular dataset's name

Step 3 - SUDO experimentation

To perform SUDO experiments, please refer to the scripts entitled train_XXX_data.py, where XXX is the particular dataset's name

Example

If you would like to conduct SUDO experiments on the Stanford DDI data, then you have to first run the code in extract_ddi_features.py and subsequently run the code in train_ddi_data.py. At present, these scripts cannot be implemented from the command line.