Delayed Impact of Fair Machine Learning

@article{Liu2018DelayedIO,
  title={Delayed Impact of Fair Machine Learning},
  author={Lydia T. Liu and Sarah Dean and Esther Rolf and Max Simchowitz and Moritz Hardt},
  journal={ArXiv},
  year={2018},
  volume={abs/1803.04383}
}
Static classification has been the predominant focus of the study of fairness in machine learning. While most models do not consider how decisions change populations over time, it is conventional wisdom that fairness criteria promote the long-term well-being of groups they aim to protect. This work studies the interaction of static fairness criteria with temporal indicators of well-being. We show a simple one-step feedback model in which common criteria do not generally promote improvement over… 

Figures from this paper

The zoo of Fairness metrics in Machine Learning
TLDR
This work tries to make some order out of this zoo of definitions of fairness in Machine Learning, that consider different notions of what is a “fair decision” in situations impacting individuals in the population.
Fairness is not static: deeper understanding of long term fairness via simulation studies
TLDR
An extensible open-source software framework for implementing fairness-focused simulation studies and further reproducible research is provided, and it is shown that static or single-step analyses do not give a complete picture of the long-term consequences of an ML-based decision system.
The Cost of Fairness: Evaluating Economic Implications of Fairness-Aware Machine Learning
TLDR
It is demonstrated that the EO is the only type of fairness-aware machine learning that can remove group-level disparity, and such fairness comes with a non-negligible cost to the firm and society.
Causal Interventions for Fairness
TLDR
This work uses causal methods to model the effects of interventions, allowing for potential interference--each individual's outcome may depend on who else receives the intervention, and demonstrates this with an example of allocating a budget of teaching resources using a dataset of schools in New York City.
Fairness-enhancing interventions in stream classification
TLDR
This work proposes fairness-enhancing interventions that modify the input data so that the outcome of any stream classifier applied to that data will be fair.
Long-Term Impacts of Fair Machine Learning
TLDR
Two feedback models describing how people react when receiving machine-aided decisions are introduced and it is illustrated that some commonly used fairness criteria can end with undesirable consequences while reinforcing discrimination.
Comparing Fairness Criteria Based on Social Outcome
TLDR
It is demonstrated that a comparison between several policies induced by well-known fairness criteria, including the color-blind (CB), the demographic parity (DP), and the equalized odds (EO), shows that the EO is the only criterion among them that removes group-level disparity.
On the Long-term Impact of Algorithmic Decision Policies: Effort Unfairness and Feature Segregation through Social Learning
TLDR
An effort-based measure of fairness is proposed and a data-driven framework for characterizing the long-term impact of algorithmic policies on reshaping the underlying population is presented.
Individual Fairness Under Composition
TLDR
It is found that fairness does not behave well under composition and proposed directions to remedy the situation are proposed.
Designing Fairly Fair Classifiers Via Economic Fairness Notions
TLDR
This paper proposes novel relaxations of these fairness notions — envy-freeness and equitability — in machine learning which apply to groups rather than individuals, and are compelling in a broad range of settings.
...
1
2
3
4
5
...

References

SHOWING 1-10 OF 34 REFERENCES
Fairness Constraints: Mechanisms for Fair Classification
TLDR
This paper introduces a flexible mechanism to design fair classifiers by leveraging a novel intuitive measure of decision boundary (un)fairness, and shows on real-world data that this mechanism allows for a fine-grained control on the degree of fairness, often at a small cost in terms of accuracy.
Inherent Trade-Offs in the Fair Determination of Risk Scores
TLDR
Some of the ways in which key notions of fairness are incompatible with each other are suggested, and hence a framework for thinking about the trade-offs between them is provided.
Avoiding Discrimination through Causal Reasoning
TLDR
This work crisply articulate why and when observational criteria fail, thus formalizing what was before a matter of opinion and put forward natural causal non-discrimination criteria and develop algorithms that satisfy them.
Fairness in Learning: Classic and Contextual Bandits
TLDR
A tight connection between fairness and the KWIK (Knows What It Knows) learning model is proved: a provably fair algorithm for the linear contextual bandit problem with a polynomial dependence on the dimension, and a worst-case exponential gap in regret between fair and non-fair learning algorithms.
Counterfactual Fairness
TLDR
This paper develops a framework for modeling fairness using tools from causal inference and demonstrates the framework on a real-world problem of fair prediction of success in law school.
Fair Inference on Outcomes
TLDR
It is argued that the existence of discrimination can be formalized in a sensible way as the presence of an effect of a sensitive covariate on the outcome along certain causal pathways, a view which generalizes (Pearl 2009).
The Myth in the Methodology: Towards a Recontextualization of Fairness in Machine Learning
TLDR
The design and adoption of machine learning tools in high-stakes social contexts should be as much a matter of democratic deliberation as of technical analysis.
Equality of Opportunity in Supervised Learning
TLDR
This work proposes a criterion for discrimination against a specified sensitive attribute in supervised learning, where the goal is to predict some target based on available features and shows how to optimally adjust any learned predictor so as to remove discrimination according to this definition.
Fair prediction with disparate impact: A study of bias in recidivism prediction instruments
TLDR
It is demonstrated that the criteria cannot all be simultaneously satisfied when recidivism prevalence differs across groups, and how disparate impact can arise when an RPI fails to satisfy the criterion of error rate balance.
On Fairness and Calibration
TLDR
It is shown that calibration is compatible only with a single error constraint, and that any algorithm that satisfies this relaxation is no better than randomizing a percentage of predictions for an existing classifier.
...
1
2
3
4
...