Conformal Prediction Sets Improve Human Decision Making

This is the codebase accompanying the paper "Conformal Prediction Sets Improve Human Decision Making", published at ICML 2024. Here we discuss how to generate the datasets of conformal prediction sets used in the paper.

Setup

The main prerequisite is to set up the python environment.

conda create --name conformal python=3.10
conda activate conformal
conda install pytorch=2.0.1 torchvision pytorch-cuda=11.7 -c pytorch -c nvidia
conda install scipy tqdm pandas
conda install transformers datasets 
conda install -c conda-forge emoji  # For go-emotions
pip install git+https://github.com/openai/CLIP.git  # For object-net
pip install span_marker  # for Few-NERD

Usage - main.py

The main script for creating datasets is unsurprisingly main.py. This script loads raw datasets, splits them, loads or trains a model, performs conformal calibration, and generates conformal prediction sets for the test set data points.

The basic usage is as follows:

python main.py --dataset <dataset>

where <dataset> is the desired dataset. We have implemented fashion-mnist, go-emotions, object-net, and few-nerd.

fashion-mnist, go-emotions, and few-nerd will be automatically downloaded when the code is run. object-net can be downloaded in its entirety from the original source and preprocessed as in this function. Since only a small subset is needed for our code, we provide a link to download only what is needed, and already pre-processed.

Dynamic Updating of Config Values

Dataset and calibration hyperparameters are loaded from the config.py file at runtime. However, it is also possible to update the hyperparameters on the command line using the flag --config. For each hyperparameter <key> that one wants to set to a new value <value>, add the following to the command line:

--config <key>=<value>

This can be done multiple times for multiple keys. A full list of config values is visible in the config.py file.

Two options for experiments

To create a fair comparison between conformal and top-k prediction sets, we ensure their sets have the same coverage on the calibration set. This can be done in two ways:

1. Default config: Select k of top-k first, and compute alpha as (1 - top_k_coverage) on the calibration set. Then use that alpha for conformal prediction.
2. Alternative: Pick an alpha value for conformal prediction first, and then select k for top-k empirically to match the coverage from conformal prediction. Since k must be an integer, we are not guaranteed to find a value of k which produces a similar alpha. Example command:

   python main.py --dataset go-emotions --config alpha=0.05 --config k=None
   

Run Directories

By default, the main command above will create a directory of the form logs/<date>_<hh>-<mm>-<ss>, e.g. Jan24_19-01-22, to store information about the run, including:

• Config files as json
• Experiment metrics / results as json
• stderr / stdout logs
• Output csvs containing conformal prediction sets for the test data

BibTeX

@inproceedings{cresswell2024conformal,
  title={Conformal Prediction Sets Improve Human Decision Making}, 
  author={Jesse C. Cresswell, Yi Sui, Bhargava Kumar, Noël Vouitsis},
  booktitle={International Conference on Machine Learning},
  year={2024}
}