Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets
Authors: Eleni Straitouri, Suhas Thejaswi, Manuel Rodriguez
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our framework using real human predictions from two different human subject studies and show that, in decision support systems based on prediction sets, there is a trade-off between accuracy and counterfactual harm. In this section, we use data from two different human subject studies to: a) evaluate the average counterfactual harm caused by decision support systems based on prediction sets; b) validate the theoretical guarantees offered by our computational framework (i.e., Corollaries 1 and 2); c) investigate the trade-off between the average counterfactual harm caused by decision support systems based on prediction sets and the average accuracy achieved by human experts using these systems. |
| Researcher Affiliation | Academia | Eleni Straitouri Max Planck Institute for Software Systems Kaiserslautern, Germany estraitouri@mpi-sws.org Suhas Thejaswi Max Planck Institute for Software Systems Kaiserslautern, Germany thejaswi@mpi-sws.org Manuel Gomez Rodriguez Max Planck Institute for Software Systems Kaiserslautern, Germany manuelgr@mpi-sws.org |
| Pseudocode | No | The paper describes computational frameworks and procedures but does not include explicit pseudocode blocks or figures labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | An open-source implementation of our methodology is publicly available at https://github.com/Networks-Learning/controlling-counterfactual-harm-prediction-sets. |
| Open Datasets | Yes | We first experiment with the Image Net16H dataset by Steyvers et al. [52], which comprises 32,431 predictions made by 145 human participants on their own about noisy images created using 1,200 unique natural images from the Image Net Large Scale Visual Recognition Challenge (ILSRVR) 2012 dataset [53]. Then, we experiment with the Image Net16H-PS dataset10 by Straitouri et al. [18], which comprises 194,407 predictions made by 2,751 human participants using decision support systems Cλ about the set of noisy images with ω = 110 described above. (Footnote 8): All classifiers and images are publicly available at https://osf.io/2ntrf. (Footnote 10): The dataset is publicly available at https://github.com/Networks-Learning/counterfactual-prediction-sets/. |
| Dataset Splits | Yes | To this end, we randomly split the images (and human predictions) into a calibration set (10%), which we use to find the harm-controlling sets Λ(α) by applying Corollary 1, and a test set (90%), which we use to estimate the average counterfactual harm H(λ) caused by the decision support systems Cλ as well as the average accuracy A(λ) of the predictions made by a human expert using Cλ. with only difference that we split the images into a calibration set (10%), which we use to find Λ(α), a validation set (10%), which we use to select the value of the hyperparameter w following a similar procedure as in Huang et al. [49], and a test set (80%). |
| Hardware Specification | Yes | All experiments ran on a Mac OS machine with an M1 processor and 16GB Memory. |
| Software Dependencies | Yes | We implement our methods and execute our experiments using Python 3.10.9, along with the open-source libraries Num Py 1.26.4 (BSD License), and Pandas 2.2.1 (BSD 3-Clause License). |
| Experiment Setup | Yes | For each stratum of images, we apply our framework to decision support systems Cλ with different pre-trained classifiers, namely VGG19 [54], Dense Net161 [55], Google Net [56], Res Net152 [57] and Alex Net [58] after 10 epochs of fine-tuning, as provided by Steyvers et al. [52]. In each iteration of the experiment, given a λ value, we select the value of w {0.02, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35} for which the set-valued predictor achieves the smallest average prediction set-size over the data samples in the validation set. |