Avoiding Discrimination through Causal Reasoning (1706.02744v2)

Abstract: Recent work on fairness in machine learning has focused on various statistical discrimination criteria and how they trade off. Most of these criteria are observational: They depend only on the joint distribution of predictor, protected attribute, features, and outcome. While convenient to work with, observational criteria have severe inherent limitations that prevent them from resolving matters of fairness conclusively. Going beyond observational criteria, we frame the problem of discrimination based on protected attributes in the language of causal reasoning. This viewpoint shifts attention from "What is the right fairness criterion?" to "What do we want to assume about the causal data generating process?" Through the lens of causality, we make several contributions. First, we crisply articulate why and when observational criteria fail, thus formalizing what was before a matter of opinion. Second, our approach exposes previously ignored subtleties and why they are fundamental to the problem. Finally, we put forward natural causal non-discrimination criteria and develop algorithms that satisfy them.

Analyzing Discrimination through Causal Reasoning in Machine Learning

The paper "Avoiding Discrimination through Causal Reasoning" addresses critical issues in fairness within the field of ML, focusing on the limitations of statistical discrimination criteria that are often observational. The authors propose a novel framework using causal reasoning to better understand and mitigate discrimination arising from protected attributes.

Overview

The paper begins by examining traditional approaches to fairness, such as demographic parity and equalized odds. These methods rely on observational data, which only considers the joint distribution of predictor~$R$, protected attribute~$A$, features~$X$, and outcome~$Y$. The authors argue that such criteria fail to address fairness comprehensively, highlighting contradictions and identifiability issues exemplified in prior research.

Contributions

The core contribution of this research lies in reframing discrimination as a causal problem, which leads to several advancements:

1. Causal Criteria for Discrimination: The authors articulate causal non-discrimination criteria that go beyond observational independence, which addresses the direct causal influence of a protected attribute that is not mitigated by any resolving variables.
2. Subtle Aspects in Fairness: Through a causal lens, the paper uncovers and formalizes subtleties in fair decision-making, particularly the distinction between inherent protected attributes and their proxies.
3. Algorithms for Causally Fair Models: The authors propose algorithms that aim to satisfy newly defined causal fairness criteria. These include methods to remove discrimination effects under certain linear assumptions about the data generation model.
4. Shift to Model-Assumption Dialogue: The work emphasizes moving from searching for statistical fairness criteria to focusing on assumptions about data generation, with causality offering a structured approach for hypothesis testing and assumption validation.

Methodology and Key Findings

The paper outlines procedures for mitigating two primary types of discrimination:

• Unresolved Discrimination: This occurs when the influence of a protected attribute on a prediction is not mitigated by known resolution paths. The authors discuss conditions in which unresolved discrimination cannot be detected purely through observation, due to the inability of observational criteria to distinguish between causally different scenarios.
• Proxy Discrimination: Defined as the unwanted influence of a proxy for a protected attribute. The authors propose a framework using causal interventions to isolate and eliminate discriminatory paths while maintaining predictive performance, supported by a linear causal model example.

Implications

The implications of adopting a causal perspective in fairness are multifaceted:

• Theoretical Robustness: The paper extends the theoretical foundation of fairness, suggesting that causal assumptions allow for a more nuanced understanding of discrimination beyond statistical independence.
• Practical Implementation: Implementing causal fairness criteria could enhance algorithmic decision-making in socially sensitive applications by incorporating domain-specific causal structures and assumptions.
• Future Research Directions: This research sets the groundwork for more sophisticated exploration into causality-based fairness metrics and algorithms, which could further refine machine learning systems' ability to make ethical decisions. Future work could explore broader classes of causal models and their application across diverse contexts.

Conclusion

By framing discrimination within causal reasoning, the paper presents a significant advancement in addressing fairness in machine learning. The proposed methods and insights encourage a deeper examination of the underlying assumptions and causal pathways in data, potentially leading to more fair and transparent predictive models. This causal approach lays the groundwork for future studies to build upon and refine fairness criteria, making substantial steps toward resolving ethical issues inherent in automated decision-making systems.