As machine-learning algorithms proliferate, there are growing concerns regarding their fairness. Can we stop AI, trained on real-world data, from reproducing and exacerbating real-world biases? There exists a growing body of work on fairness in AI, but often focussing rather narrowly on classification problems. Online advertising and job-candidate rankings, for example, utilize ranking algorithms instead of classification algorithms. The objectives of this work are to (a) describe the theoretical basis of fairness in ranking and the metrics used to evaluate it, (b) explore several existing post-processing algorithms for fairness in ranking, (c) introduce a novel randomized algorithm based on Mallows distribution that offers a tradeoff between fairness and accuracy, and (d) compare its performance in terms of obtained fairness and loss of ranking accuracy to such of existing deterministic algorithms.
As machine-learning algorithms proliferate, there are growing concerns regarding their fairness. Can we stop AI, trained on real-world data, from reproducing and exacerbating real-world biases? There exists a growing body of work on fairness in AI, but often focussing rather narrowly on classification problems. Online advertising and job-candidate rankings, for example, utilize ranking algorithms instead of classification algorithms. The objectives of this work are to (a) describe the theoretical basis of fairness in ranking and the metrics used to evaluate it, (b) explore several existing post-processing algorithms for fairness in ranking, (c) introduce a novel randomized algorithm based on Mallows distribution that offers a tradeoff between fairness and accuracy, and (d) compare its performance in terms of obtained fairness and loss of ranking accuracy to such of existing deterministic algorithms.
Fairness in Rankings via Randomized Algorithms
KLIACHKIN, ANDRII
2023/2024
Abstract
As machine-learning algorithms proliferate, there are growing concerns regarding their fairness. Can we stop AI, trained on real-world data, from reproducing and exacerbating real-world biases? There exists a growing body of work on fairness in AI, but often focussing rather narrowly on classification problems. Online advertising and job-candidate rankings, for example, utilize ranking algorithms instead of classification algorithms. The objectives of this work are to (a) describe the theoretical basis of fairness in ranking and the metrics used to evaluate it, (b) explore several existing post-processing algorithms for fairness in ranking, (c) introduce a novel randomized algorithm based on Mallows distribution that offers a tradeoff between fairness and accuracy, and (d) compare its performance in terms of obtained fairness and loss of ranking accuracy to such of existing deterministic algorithms.File | Dimensione | Formato | |
---|---|---|---|
dissertation_kliachkin.pdf
accesso aperto
Dimensione
1.07 MB
Formato
Adobe PDF
|
1.07 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/64789