Data and Algorithm Ethics

Status: concluded
Period: November 2017 – November 2020
Funding: 69500 €
Funding organization: Fondazione Bruno Kessler
Person(s) in charge: Elena Beretta

The main goal of this research is to investigate the impact of data revolution in our society, combining data analytics techniques with an epistemological approach from social sciences and humanities.

We are living in an historical age in which automated decision making are rapidly spreading in nowadays economies transforming several domains of our daily life. Over the last few decades, such tools have become progressively complex by exploiting an ever-growing amount of personal and proxy data through increasingly sophisticated massive profiling techniques. Especially, ranking systems are one of the most relevant technologies in the transformation we are witnessing and there is almost nothing today that is not expressed in a ranking form: people, products, jobs, opinions. Ranking is therefore one of the predominant forms by which both online and offline software systems present results in a wide variety of domains ranging from web search engines to recommendation systems. Despite they have been widely employed since decades in Information Retrieval domain, they have recently come back at the cutting edge thanks to the explosive growth of computational power and data availability. The main task of ranking systems is to find an allocation of elements to each of the n positions so that the total value obtained is maximized. The success of an element and its exposure are strongly dependent by its position; a number of recent researches have shown that ranking algorithms whose only task is to maximize utility do not necessarily lead to fair or desirable scenarios, causing forms of algorithmic biases that in some cases can lead to serious social implications. Web search engine results that inadvertently promote stereotypes through over-representation of sensitive attributes such as gender, ethnicity and age are an example. For this reason, the study of fairness and bias in ranking algorithms is a fundamental task in this transformation era.

In recent years, several formal definitions of fairness have been suggested by the automated-system community, especially by the machine learning one. Despite the increasing influence of rankings on our society and economy, fairness in ranking systems is still a poorly explored ground. The aim of the research is to contribute to the debate on fairness in machine learning, and in particular in ranking systems domain. Main goals are briefly summarized below: i) discussing the impact that recent proposed fairness formalizations may have on societies; ii) proposing new solutions to recalibrates ranking outcomes subject to fairness constraints; iii) studying the trade-off among fairness and traditional ranking systems utility.

The main achievements of this PhD research project are briefly resumed:
Journal Papers:
(i) Antonio Vetrò, Antonio Santangelo, Elena Beretta, Juan Carlos De Martin. (2019) AI: from rational agents to socially responsible agents, Digital Policy, Regulation and Governance, https://doi.org/10.1108/DPRG-08-2018-0049
Conference Proceedings:
(ii) Elena Beretta, Antonio Santangelo, Antonio Vetrò, Bruno Lepri, Juan Carlos De Martin. (2019) The Invisible Power of Fairness. How Machine Learning Shapes Democracy. In: Meurs MJ., Rudzicz F. (eds) Advances in Artificial Intelligence. Canadian AI 2019. Lecture Notes in Computer Science, vol 11489. Springer, Cham. https://doi.org/10.1007/978-3-030-18305-9_19
(iii) Elena Beretta, Antonio Vetrò, Bruno Lepri, Juan Carlos De Martin. (2018) Ethical and Socially Aware Data Labels, Information Management and Big Data. SIM-Big 2018. Communications in Computer and Information Science, vol 898, pp. 320-327, Springer, Cham. https://doi.org/10.1007/978-3-030-11680-4_30
Working Paper:
Nexa Center, Fairness e Machine Learning, December 2018, Working paper nr 2/2018.
Accepted for publication:
(i) Elena Beretta, Antonio Vetrò, Bruno Lepri, Juan Carlos De Martin. (2021) Equality of opportunity in ranking: a Fair-Distributive Model, Second International Workshop on Algorithmic Bias in Search and Recommendation (April 2021)

@conference{<LineBreak>	11583_2898812,

title = {Equality of Opportunity in Ranking: a Fair-Distributive Model},

author = {Elena Beretta and Antonio Vetrò and Bruno Lepri and Juan Carlos De Martin},

url = {https://link.springer.com/chapter/10.1007/978-3-030-78818-6_6#citeas},

doi = {10.1007/978-3-030-78818-6_6},

isbn = {978-3-030-78817-9},

year  = {2021},

date = {2021-01-01},

urldate = {2021-01-01},

booktitle = {Advances in Bias and Social Aspects in Search and Recommendation},

pages = {51–63},

publisher = {Springer},

abstract = {In this work, we define a Fair-Distributive ranking system based on Equality of Opportunity theory and fair division models. The aim is to determine the ranking order of a set of candidates maximizing utility bound to a fairness constraint. Our model extends the notion of protected attributes to a pool of individual circumstances, which determine the membership to a specific type. The contribution of this paper are i) a Fair-Distributive Ranking System based on criteria derived from distributive justice theory and its applications in both economic and social sciences; ii) a class of fairness metrics for ranking systems based on the Equality of Opportunity theory. We test our approach on a hypothetical scenario of a selection university process. A follow-up analysis shows that the Fair-Distributive Ranking System preserves an equal exposure level for both minority and majority groups, providing a minimal system utility cost.},

keywords = {},

pubstate = {published},

tppubtype = {conference}

}

@inproceedings{beretta2021detecting,

title = {Detecting discriminatory risk through data annotation based on bayesian inferences},

author = {Elena Beretta and Antonio Vetrò and Bruno Lepri and Juan Carlos De Martin},

url = {https://nexa.polito.it/wp-content/uploads/2024/06/2021-facct-risks.pdf},

year  = {2021},

date = {2021-01-01},

urldate = {2021-01-01},

booktitle = {Proceedings of the 2021 ACM conference on fairness, accountability, and transparency},

pages = {794–804},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}