This is a repository containing the code used in the paper Improving Expert Predictions with Prediction Sets, ICML 2023.

Experiments ran on python 3.9.2, with GPU. To install the required libraries, set the torch version according to the available device (CPU or GPU) in requirements.txt and run:

pip install -r requirements.txt

conformal_prediction.py contains functions to implement and evaluate our conformal prediction based decision support system, with the following key components:

• find_all_alpha_values finds all values of $\alpha$ to be examined, using the calibration set.
• find_alpha_star finds the near optimal $\alpha$ value, based on the estimation of the expert's success probability (works for standard and modified conformal prediction).

• config.py contains the configuration details for both the synthetic and real data experiments.
• utils.py contains functions for generating/reading the synthetic and real datasets.
• ./model/model.py contains the classes for the models used in the synthetic and real data experiments.
• ./expert/expert.py contains the classes for the experts used in the synthetic and real data experiments.

Under scripts/single/{synthetic|real} are scripts for synthetic and real data experiments given the calibration data split, and in case of synthetic data, additionally given the label space size.

Scripts for both synthetic and real data:

• standard_cp.py runs experiments using standard conformal prediction, given the calibration and estimation set splits and the number of each experiment's repetitions using random splits. In case of synthetic data, it takes additionally as argument the label space size, and runs experiments for all combinations of experts and models with success probabilities $\in{0.3,0.5,0.7,0.9}$. In case of real data, it runs experiments for all the pre-trained models.
• avg_size.py computes and stores the average prediction set size for each value of $\alpha$. It takes the same arguments and runs similarly to standard_cp.py.
• coverage.py computes the near optimal alpha value and then measures the empirical coverage during test. It takes the same arguments as standard_cp.py.
• topk.py measures the empirical expert misprediction probability during test using a top $k$ predictor. It takes the same arguments as standard_cp.py and additionally takes the $k$ value. It runs similarly as standard_cp.py.
• robustness_analysis.py measures the empirical expert misprediction probability during test under IIA violations as described in Appendix E. It takes the same arguments and runs similarly to standard_cp.py.

Scripts only for synthetic data:

• modified_cp.py runs experiments using modified conformal prediction. It takes the same arguments and runs similarly to standard_cp.py.
• set_size_distribution.py computes the subset size distribution for the near optimal $\alpha$ value. It takes the same arguments and runs similarly to scripts/single/synthetic/standard_cp.py.

Scripts only for real data:

• more_expressive.py runs the experiments experiments using a discrete choice model with a more expressive context including the difficulty level of samples. It takes the same arguments and runs similarly to scripts/single/real/standard_cp.py.

Under scripts/batch/{synthetic|real} are scripts for running multiple synthetic and real data experiments for several calibration data splits, and in case of synthetic data, for several label space sizes. scripts/batch/synthetic/run_all_splits_nlabels.py runs scripts/single/synthetic/standard_cp.py, scripts/batch/real/run_all_splits.py runs scripts/single/real/standard_cp.py, and each run_all_<script>.py runs the relevant <script>.

Functions to produce plots and numerical results are in the below scripts. More specifically:

• ./plot/plot.py contains all plotters to produce the figures appeared in the paper.
• ./plot/printers.py contains all the functions that compute all the numerical results reported in the paper.
• plots.ipynb produces the figures appeared in the paper in presence of the results.

For synthetic data experiments run:

python3 -m scripts.single.synthetic.standard_cp --n_labels <n> --cal_split <split> --runs <runs>

where:

• <n> is the total number of labels.
• <split> is the calibration and estimation split.
• <runs> is the number of times that each experiment will run with different random splits of the above specified size.

Results will be stored under ./results_synthetic. To run all experiments for <n> $\in{10,50,100}$, <split> $\in {0.02, 0.05, 0.1, 0.15,}$ and <runs> $=10$, run:

python3 ./scripts/batch/synthetic/run_all_splits_nlabels.py

For real data experiments run (<split> and <runs> same as above):

python3 -m scripts.single.real.standard_cp --cal_split <split> --runs <runs>

Results will be stored under ./results_real. To run all experiments for <split> $\in {0.02, 0.05, 0.1, 0.15,}$ and <runs> $=10$, run:

python3 ./scripts/batch/real/run_all_splits.py

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

@inproceedings{straitouri23improving,
         title={Improving Expert Predictions with Prediction Sets},
         author={Straitouri, Eleni and Wang, Lequn and Okati, Nastaran and Gomez-Rodriguez, Manuel},
         booktitle={Proceedings of the 40th International Conference on Machine Learning},
         year={2023}
         
}