Causal Context Connects Counterfactual Fairness to Robust Prediction and Group Fairness

Authors: Jacy Anthis, Victor Veitch

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we conduct brief experiments in a semi-synthetic setting with the Adult income dataset [3] to confirm that a counterfactually fair predictor under these conditions achieves out-of-distribution accuracy and the corresponding group fairness metric.
Researcher Affiliation Academia 1University of Chicago, 2University of California, Berkeley, 3Sentience Institute
Pseudocode No The paper defines mathematical theorems and a predictor (Theorem 3) but does not present a structured pseudocode block or algorithm.
Open Source Code Yes code to reproduce these results or produce results with varied inputs (number of datasets sampled, effect of A on X, probabilities of each bias, type of predictor) is available at https://github.com/jacyanthis/Causal-Context.
Open Datasets Yes We used the Adult income dataset [3] with a simulated protected class A, balanced with P(A = 0) = P(A = 1) = 0.5. [3] refers to 'Barry Becker and Ronny Kohavi. Adult. 1996. DOI: 10 . 24432 / C5XW20.'
Dataset Splits No The paper mentions using the Adult income dataset and training predictors but does not specify how the dataset was split into training, validation, and test sets (e.g., percentages or sample counts).
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU specifications, or cloud computing instance types used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes On each dataset, we trained three predictors: a naive predictor trained on A and X, a fairness through unawareness (FTU) predictor trained only on X, and a counterfactually fair predictor based on an average of the naive prediction under the assumption that A = 1 and the naive prediction under the assumption A = 0, weighted by the proportion of each group in the target distribution.