Research map / Probabilistic and scientific modeling / Causal discovery

Local Causal Discovery for Structural Evidence of Direct Discrimination

AAAI 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LD3, a local causal discovery method that finds parents of an outcome in a linear number of CI tests to yield structural evidence and a valid adjustment set for the weighted controlled direct effect, applied to COMPAS recidivism and liver transplant allocation.

Problem

Identifying direct discrimination in causal fairness analysis when the causal graph is unknown, where global discovery is unstable and costly on finite samples.

Approach

Local discovery that labels variables by causal partition relative to exposure and outcome to recover the parents of the outcome with O(|Z|) CI tests, then uses those parents as a valid adjustment set for the WCDE and tests the SDC via d-separation.

Key results

• On Sangiovese synthetic benchmark, LD3 was generally most performant across metrics, with LDP performing similarly; NOTEARS was significantly slower than other methods.
• On COMPAS, LD3 stably predicted the WCDE of race on two-year recidivism as not significant (p = 0.87, 0.68, 0.77), while PC and LDECC fluctuated between p=0.000 and p>0.4; PC and LDECC took 46x longer and performed at least 11.7x more tests.
• On liver transplant data, PC performed 479-1021x more tests and took 2295-5870x longer than LD3; LD3 detected direct discrimination (SDC=1) with significant WCDE (p<0.005) in all settings.

Measured effects reported in the paper

Surprising findings

• PC returned untrustworthy edges such as weight to sex and height to sex, and LDECC predicted the outcome had children, which is known to be untrue.
• LD3 predicted juvenile delinquency as a parent of COMPAS decile score at all significance levels, while PC and LDECC never did.

Limitations

• Assumes Y has no descendants among observed variables, so cannot handle outcomes with descendants.
• Cannot differentiate confounder-type from mediator-type parents (Z1 vs Z3), limiting analysis of indirect and spurious discrimination.
• Zero WCDE does not guarantee absence of direct discrimination; results are preliminary qualitative indicators.

Experimental setup

Datasets: COMPAS (ProPublica) STAR (UNOS liver transplant data) Sangiovese (bnlearn) Asia (bnlearn) Sachs (bnlearn)

Benchmarks: Sangiovese Asia Sachs COMPAS STAR liver transplant

Metrics: parent F1 precision recall runtime number of CI tests WCDE estimate and p-value SDC accuracy estimator variance

Baselines: PC LDECC MB-by-MB LDP DirectLiNGAM NOTEARS

Builds on: Local Discovery by Partitioning (LDP) causal partition taxonomy double machine learning PC algorithm

Closest papers by meaning

• FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.89)
• Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.88)
• A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs (NeurIPS 2024 · similarity 0.88)
• Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.88)
• Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.88)
• Identifying Structural Biases from Causal Mechanism Shifts (NeurIPS 2026 · similarity 0.87)