Fairness seen as global sensitivity analysis

Article Statistics & Probability

A Survey of Bias in Machine Learning Through the Prism of Statistical Parity

Philippe Besse et al.

Summary: This article discusses the introduction of bias in machine learning models and methods to measure and reduce such bias. Testing different approaches on the adult income dataset reveals that creating fair machine learning models can be a challenging task.

AMERICAN STATISTICIAN (2022)

添加到收藏夹

Article Statistics & Probability

FOUNDATIONS OF STRUCTURAL CAUSAL MODELS WITH CYCLES AND LATENT VARIABLES

Stephan Bongers et al.

Summary: Structural causal models (SCMs) are commonly used for causal modeling, with acyclic SCMs forming a subclass that allows for latent confounders. This paper explores SCMs in a more general setting, showing that properties of acyclic SCMs may not hold in the presence of cycles. Some properties, such as unique distributions and Markov property, can be maintained for SCMs under certain solvability conditions. The introduction of simple SCMs extends the convenience of acyclic SCMs to models with cycles.

ANNALS OF STATISTICS (2021)

添加到收藏夹

Proceedings Paper Computer Science, Artificial Intelligence

Fairness by Explicability and Adversarial SHAP Learning

James M. Hickey et al.

Summary: The study introduces a new definition of fairness that emphasizes the role of an external auditor and model explainability, and provides a regularization framework constructed from SHAP values to mitigate model bias. By linking fairness explainability constraints to classical statistical fairness metrics for binary classification tasks, the study demonstrates the performance of fair models on synthetic dataset, UCI Adult dataset, and a real-world credit scoring dataset.

MACHINE LEARNING AND KNOWLEDGE DISCOVERY IN DATABASES, ECML PKDD 2020, PT III (2021)

添加到收藏夹

Proceedings Paper Computer Science, Information Systems