|
Title
|
|
|
|
Explainability methods to detect and measure discrimination in machine learning models
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| Today, it is common to use machine learning models for high-stakes decisions, but this can pose a threat to fairness as these models can amplify bias present in the dataset. At the moment, there is no consensus on a universal method to tackle this, and we argue that this is also not possible as the right method will depend on the context of each case. As a solution, our aim was to bring transparency in the fairness domain, and in earlier work, we proposed a counterfactual-based algorithm (ππππΆππΉ) to identify bias in machine learning models. This method attempts to counter the disagreement problem in Explainable AI, by reducing the flexibility of the model owner. We envision a future where transparency tools such as the latter are used to perform fairness audits by independent auditors who can judge for each case whether the audit revealed discriminatory patterns or not. This approach would be more in line with the current nature of EU legislation, as its requirements are often too contextual and open to judicial interpretation to be automated. |
| |
|
Language
|
|
|
|
English
| |
|
Source (journal)
|
|
|
|
CEUR workshop proceedings
| |
|
Source (book)
|
|
|
|
EWAFβ23 : European Workshop on Algorithmic Fairness, June 07β09, 2023, Winterthur, Switzerland
| |
|
Publication
|
|
|
|
CEUR
,
2023
| |
|
Volume/pages
|
|
|
|
3442
, p. 1-5
| |
|
Full text (open access)
|
|
|
|
| |
|