Counterfactual Fairness with Partially Known Causal Graph

Authors: Aoqi Zuo, Susan Wei, Tongliang Liu, Bo Han, Kun Zhang, Mingming Gong

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results on both simulated and real-world datasets demonstrate the effectiveness of our method. In this section, we illustrate our approach on a simulated and a real-world dataset by evaluating the prediction performance and fairness of our approach.
Researcher Affiliation Academia Aoqi Zuo The University of Melbourne azuo@student.unimelb.edu.au Susan Wei The University of Melbourne susan.wei@unimelb.edu.au Tongliang Liu The University of Sydney tongliang.liu@sydney.edu.au Bo Han Hong Kong Baptist University bhanml@comp.hkbu.edu.hk Kun Zhang Carnegie Mellon University & MBZUAI kunz1@cmu.edu Mingming Gong The University of Melbourne mingming.gong@unimelb.edu.au
Pseudocode Yes Algorithm 1 Finding the critical set of S with respect to T in an MPDAG ... Algorithm 2 Identify the type of ancestral relation of S with respect to T in an MPDAG
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes]
Open Datasets Yes The UCI Student Performance Data Set [10] regarding students performance in Mathematics is used in this experiment.
Dataset Splits No The proportion of training and test data is splitted as 0.8 0.2.
Hardware Specification No The paper does not provide specific details about the hardware used to run its experiments. The self-assessment section indicates N/A for compute resources.
Software Dependencies No The paper mentions tools like "GES structure learning algorithm [5]" and "TETRAD [45]" but does not specify their version numbers or other ancillary software with specific versions.
Experiment Setup Yes The synthetic data is generated from linear structural equation models according to a random DAG. ...The sensitive attribute can have two or three values, drawn from a Binomial([0,1]) or Multinomial([0,1,2]) distribution separately. The weight, βij, of each directed edges Xi Xj in the generated DAG, is drawn from a Uniform([ 2, 0.5] < [0.5, 2]) distribution. The data are generated according to the following linear structural equation model: ...where 1, ..., n are independent N(0, 1.5). Then we generate one sample with size 1000 for each DAG. ...fitting a linear regression model