Fairness in link analysis ranking algorithms
Keywords: fairness, link analysis ranking, hits, pagerank, social networks, evolving network models, information retrieval, search algorithms
TL;DR: We provide theoretical and empirical evidence for when link analysis ranking algorithms are unfair towards minority groups, and we propose methods for improving the ranking of minorities using network structures.
Abstract: In this paper, we investigate the conditions under which minority groups get underrepresented (suppressed) in rankings produced by link analysis ranking algorithms, leading to biased rankings. As recent work shows that link analysis algorithms often prevent minority groups from reaching high rankings, we take a step further in analyzing when do such algorithms amplify pre-existing bias and when can they alleviate it. We find that the most common algorithms using link analysis to create rankings based on nodes' centralities, such as Pagerank and HITS, produce vastly different outcomes: compared to the bias encoded in the degree distribution of a network with multiple communities, Pagerank often mirrors the degree distribution for most of the ranking positions and it can equalize representation of minorities among the top ranked nodes; on the other hand, we find that HITS amplifies pre-existing bias in homophilic networks through a novel theoretical analysis. We find the root cause of bias amplification to be the level of homophily, as well as inequality in the degree distribution. We characterize fundamental differences in how common algorithms may be affected by bias, and explore a series of algorithmic variations in the search for fairness. We find that randomization is a promising tool in debiasing deep inequities encoded in link structures. This work paves the way towards a deep understanding on the difficulty of fixing feature bias in ranking, as the scores that link analysis algorithms output are often used as features in learning-to-rank algorithms, implying that biased features will have a lasting effect on the fairness of many ranking schemes. We illustrate our theoretical analysis on both synthetic and real datasets.
Track: Social Networks, Social Media, and Society
Submission Guidelines Scope: Yes
Submission Guidelines Blind: Yes
Submission Guidelines Format: Yes
Submission Guidelines Limit: Yes
Submission Guidelines Authorship: Yes
Student Author: No
Submission Number: 1784
Loading