Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets

This repository contains the code used in the paper "Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets", accepted at the ICML 2024 Workshop on Humans, Algorithmic Decision-Making and Society and published at NeurIPS 2024.

Install Dependencies

Experiments ran on python 3.10.9. To install the required libraries run:

pip install -r requirements.txt

Code Structure

Data

The folder data/ includes the ImageNet16H dataset and the data/study_pred_sets includes the ImageNet16H-PS dataset. It also includes the following scripts to pre-process the datasets for the experiments:

• make_strata_per_noise.py stratifies the images, classifier and human predictions for each amount of noise $\omega\in {80,95,110}$ and stores the stratified datasets under /data. It also computes for each pair of images and prediction sets (using both vanilla and SAPS set-valued predictors) the predicted average accuracy of human experts using the mixture of MNL models.
• utils.py includes a helper function that stratifies images with the same amount of noise into two different levels of difficulty.

Setup

• config.py includes all configuration parameters of the experimental setup.
• utils.py implements helper functions for saving results and for computing the SAPS scores.
• counterfactual_harm.py implements the main module to control and evaluate the counterfactual harm (bound) of set-valued predictors (either using vanilla or SAPS non-conformity scores) under the counterfactual and interventional monotonicity assumptions for a given classifier, noise level and calibration set.

Scripts for Running Experiments and Evaluation

The folder scripts/ includes the following scripts to execute the experiments:

• counterfactual_harm.py (control mode) computes for each control level $\alpha$ from 0 to 1 with step $0.01$ the harm controlling $\lambda$ values given by Corollary 1 and all controlling $\lambda$ values that are $\leq \check{\lambda}(\alpha)$ given by Thm 2 (combining these values with the ones of Corollary 1 we have the harm controlling values of Corollary 2), for 50 random samplings of the test and calibration set, given a fixed classifier and noise level.
• counterfactual_harm.py (test mode) computes the average counterfactual harm (bound) for each $\lambda$ value across the 50 random samplings of the test set for a given classifier and noise level.
• control_all_models_noises.py executes scripts/counterfactual_harm.py for all classifiers and noise levels under control mode.
• test_all_models_noises.py executes scripts/counterfactual_harm.py for all classifiers and noise levels under test mode.
• test_metrics.py estimates the average accuracy, average prediction set size and empirical coverage per $\lambda$ value for each one of the 50 random samplings of the test set for a given classifier and noise level.
• metrics_all_models_noises.py executes scripts/test_metrics.py for each classifier and noise level.

Plots

plotters/ includes the following scripts to produce the plots in the paper:

• interventional.py produces the plots related to the average accuracy per prediction set size for images of similar difficulty across all experts and across experts with the same level of competence for each ground truth label.
• harm_vs_metric.py produces the plots of average counterfactual harm (bound) against average accuracy for all noise levels and classifiers as well as the plots of average counterfactual harm against average prediction set size and against empirical coverage.

Running Instructions

Run Experiments

To create the stratified datasets per noise level run:

python -m data.make_strata_per_noise

To compute all controlling $\lambda$ values (for each control level $\alpha$) for each of the 50 samplings of the calibration set for each classifier and noise level run:

python -m scripts.control_all_models_noises

To evaluate the average counterfactual harm (bound) for each $\lambda$ value for each one of the 50 samplings of the test set, and every classifier and noise level run:

python -m scripts.test_all_models_noises

To compute the average accuracy (predicted using the mixture of MNLs and real using ImageNet16H-PS dataset) for each $\lambda$ value for each one of the 50 samplings of the test set, and every classifier and noise level run:

python -m scripts.metrics_all_models_noises

All results will be saved under the folder results. The above experiments use by default set-valued predictors using the vanilla non-conformity scores. To repeat the experiments using the set-valued predictors using the SAPS scores follow the instructions in the scripts above.

Plots

To produce the plots of average counterfactual harm against average accuracy for all noise levels and classifiers as well as the plots of average counterfactual harm against average prediction set size and against empirical coverage run

python -m plotters.harm_vs_metric

To produce the plots related to the average accuracy per prediction set size for images of similar difficulty across all experts and across experts with the same level of competence for each ground truth label run:

python -m plotters.interventional

All plots will be saved under the folder plots.

Citation

If you use parts of the code/data in this repository for your own research purposes, please consider citing:

@inproceedings{straitouri2024controlling,
  title={Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets},
  author={Straitouri, Eleni and Thejaswi, Suhas and Rodriguez, Manuel Gomez},
  booktitle={Advances in Neural Information Processing Systems},
  year={2024}
}