Fairness Rising from the Ranks: HITS and PageRank on Homophilic Networks (2402.13787v3)

Abstract: In this paper, we investigate the conditions under which link analysis algorithms prevent minority groups from reaching high ranking slots. We find that the most common link-based algorithms using centrality metrics, such as PageRank and HITS, can reproduce and even amplify bias against minority groups in networks. Yet, their behavior differs: one one hand, we empirically show that PageRank 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, supported by empirical results. We find the root cause of bias amplification in HITS to be the level of homophily present in the network, modeled through an evolving network model with two communities. We illustrate our theoretical analysis on both synthetic and real datasets and we present directions for future work.

• The paper demonstrates that PageRank reduces bias among top-ranked nodes, though its fairness diminishes in deeper network layers.
• The paper finds that HITS amplifies inherent biases by reinforcing densely interconnected communities and limiting minority visibility.
• The paper explores randomization and subspace HITS methods as promising strategies to mitigate bias and enhance algorithmic fairness.

Exploring the Impact of Link Analysis Algorithms on Minority Representation in Homophilic Networks

Understanding the Dynamics of PageRank and HITS

Recent work explores how link analysis ranking algorithms, notably PageRank and Hyperlink-Induced Topic Search (HITS), impact minority representation in social and information networks. These algorithms are fundamental in determining the visibility and ranking of nodes in a network, directly affecting which voices are heard and which remain obscure.

Dissecting the Bias Amplification

The paper highlights that while both PageRank and HITS are predicated on network structure, they mediate information dissemination differently. PageRank, a metric of global node centrality, can potentially correct for bias at the upper echelons of rank. However, its corrective ability does not uniformly extend through all network layers, and beyond the top nodes, it mirrors the inherent bias present in node connectivity.

Conversely, HITS, which distinguishes between 'hubs' and 'authorities,' tends to amplify existing biases, especially in homophilic networks. The bifurcation of roles in HITS creates a feedback loop that disproportionately benefits nodes within densely interconnected communities, thus reinforcing majority dominance and minority invisibility.

The paper introduces a rigorous analytical framework that isolates the conditions under which these divergent outcomes manifest. Through both theoretical analysis and empirical tests on synthetic and real-world datasets, it confirms that HITS's bias amplification is particularly pronounced in networks exhibiting strong homophily.

Introducing Randomization and Subspace HITS

A notable direction explored for mitigating bias is the introduction of randomization into the HITS algorithm. Randomized HITS, which incorporates random restarts similar to PageRank, shows promise in mirroring, and in some cases, improving upon the fairness of the degree ranking baseline. This suggests that randomness offers a buffer against the self-reinforcing mechanisms of bias.

Further, the paper investigates the Subspace HITS variation, utilizing multiple eigenvectors to compute node authority scores. While presenting an innovative method to potentially enhance fairness, the results underscore the complexity of choosing an 'optimal' number of dimensions. Different datasets yield varying fairness outcomes depending on the number of eigenvectors involved.

Forward Path

This research builds a crucial understanding of how structural properties of networks and the design of ranking algorithms intersect to shape the visibility of minority groups. It points to substantial future work in fine-tuning algorithmic interventions that could promote fairness. A particularly intriguing avenue is the systemic exploration of how network evolution, in response to altered ranking strategies, influences long-term fairness and representation equity.

Concluding Remarks

The findings underscore a nuanced landscape where algorithm design, network structure, and societal dynamics collide. The work lays down a gauntlet for future research to further unravel the multifaceted relationship between algorithmic fairness and network topology, with the hope of fostering inclusivity in digital spaces.