Disentangling Influence

This repository contains the code from our 2019 paper on indirect influence audits. We leverage disentangled representations to capture indirect influence on both individual and aggregate model predictions. Please see our paper for an explanation of our methods.

Above: Direct (left) and indirect (right) influence distributions in a synthetic example where a model learns to predict label x+y given independent features x, y, and c. Unlike direct influence audits, indirect influence captures that features such as 2x and 2y influence the prediction as proxy variables for x and y.

Installation

• If you don't already have python3, you can install the current version from here.
• Navigate to the directory where you would like to download the package and run the command:

$ git clone https://github.com/charliemarx/disentangling-influence

• Navigate to the top directory and run the following commands:

$ python setup.py
$ pip install -r requirements.txt

• To run experiments with the dSprites dataset you will need to clone the dSprites dataset github repository into DisentanglingInfluence/data/dsprites/dsprites-dataset. You can do this by navigating to DisentanglingInfluence/data/dsprites/ then running the command:

$ git clone https://github.com/deepmind/dsprites-dataset

Usage

Scripts to run our experiments are in the DisentanglingInfluence/scripts directory. Output files are saved to the outputs/<experiment_name>_test directory.