FairPFN: A Tabular Foundation Model for Causal Fairness

Machine learning (ML) systems are utilized in critical sectors such as healthcare, law enforcement, and finance, but often rely on historical data that contains demographic biases, leading to decisions that perpetuate or intensify existing inequalities. Causal and counterfactual fairness provide a transparent, human-in-the-loop framework to mitigate algorithmic discrimination, aligning closely with legal doctrines of direct and indirect discrimination. However, current causal fairness frameworks hold a key limitation in that they assume prior knowledge of the correct causal model, restricting their applicability in complex fairness scenarios where causal models are unknown or difficult to identify. To bridge this gap, we propose FairPFN, a tabular foundation model pre-trained on synthetic causal fairness data to identify and mitigate the causal effects of protected attributes in its predictions. FairPFN's key contribution is that it requires no knowledge of the causal model and demonstrates strong performance across a diverse set of hand-crafted and real-world causal scenarios relative to robust baseline methods. FairPFN paves the way for a promising direction for future research, making causal fairness more accessible to a wider variety of complex fairness problems.

When a machine learning algorithm used to make critical decisions is trained on historical data containing evidence of discrimination, these algorithms are well known to reproduce, and even worsen these ethnic, gender, or age-based biases in their predictions. For example, a hiring algorithm trained on past hiring decisions may learn to discriminate against potential candidates on account of their gender, ethnicity, or physical ability. The research field focused on detecting and mitigating these biases in machine learning algorithms is called algorithmic fairness. While many past attempts to mitigate algorithmic biases focused on "outcome-based" fairness, similar in essence to Affirmative Action or quota strategies, recent approaches focus on first understanding the source of discrimination, and then coming up with a fair alternative decision making process. One way of understanding the mechanisms of discrimination is through so-called causal graphs and models, which depict a network describing the cause and effect relationships between variables. While process-based fairness offers numerous advantages over its outcome-based alternative, including a strong analogy to US and EU anti-discrimination law, coming up with a plausable causal model for a real-world scenario is often very challenging. In order to be able to perform "causal fairness" without strict knowledge of these causal mechanisms, we leverage recent advancements in machine learning called prior-data-fitted networks (PFNs), which allow us to simultaneously identify the mechanisms of discrimination for a given situation and perform the appropriate adjustment to achieve a fair decision making process. Our method, which we call FairPFN, is able to do this from only observational data, and thus makes causal fairness increasingly accessible to machine learning practitioners.