Technologies based on data are frequently adopted in many sensitive environments to build models that support important and life-changing decisions. As a result, for an application to be ethically reliable, it should be associated with tools to discover and mitigate bias in data, in order to avoid (possibly unintentional) unethical behaviors and the associated consequences. In this paper we propose a novel solution that, exploiting the notion of Functional Dependency and its variants - well-known data constraints - aims at enforcing fairness by discovering and solving discrimination in datasets. Our system first identifies the attributes of a dataset that encompass discrimination (e.g. gender, ethnicity or religion), generating a list of dependencies, then, based on this information, determines the smallest set of tuples that must be added or removed to mitigate such bias in the dataset. Experimental results on two real-world datasets demonstrated that our approach can greatly improve the ethical quality of data sources.

Functional Dependencies to Mitigate Data Bias

Azzalini F.;Criscuolo C.;Tanca L.
2022-01-01

Abstract

Technologies based on data are frequently adopted in many sensitive environments to build models that support important and life-changing decisions. As a result, for an application to be ethically reliable, it should be associated with tools to discover and mitigate bias in data, in order to avoid (possibly unintentional) unethical behaviors and the associated consequences. In this paper we propose a novel solution that, exploiting the notion of Functional Dependency and its variants - well-known data constraints - aims at enforcing fairness by discovering and solving discrimination in datasets. Our system first identifies the attributes of a dataset that encompass discrimination (e.g. gender, ethnicity or religion), generating a list of dependencies, then, based on this information, determines the smallest set of tuples that must be added or removed to mitigate such bias in the dataset. Experimental results on two real-world datasets demonstrated that our approach can greatly improve the ethical quality of data sources.
2022
Data Bias
Fairness
Functional Dependencies
File in questo prodotto:
File Dimensione Formato  
paper66.pdf

Accesso riservato

: Publisher’s version
Dimensione 278.09 kB
Formato Adobe PDF
278.09 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1231770
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? ND
social impact