Recommendation Fairness in Social Networks Over Time

TL;DR

This work addresses fairness in social recommendations within evolving networks by analyzing three real-world dynamic datasets (Enron, Norwegian, DBLP) and evaluating six recommender methods over time. It introduces a temporal framework with metrics like $VD$, $sVD$, and $rVD$, and links fairness to dynamic properties such as minority ratio $MR$, homophily ratio $HR$, and edge densities. The study finds that overall fairness improves over time across methods, with robust associations between $MR$, $HR$, and fairness; extreme values in $HR$ can destabilize fairness in counterfactual scenarios. The Bernoulli-Barabasi-Albert counterfactual experiments provide policy-relevant insights, suggesting that targeted structural interventions can promote long-term fairness while avoiding adverse effects on utility or ranking fairness.

Abstract

In social recommender systems, it is crucial that the recommendation models provide equitable visibility for different demographic groups, such as gender or race. Most existing research has addressed this problem by only studying individual static snapshots of networks that typically change over time. To address this gap, we study the evolution of recommendation fairness over time and its relation to dynamic network properties. We examine three real-world dynamic networks by evaluating the fairness of six recommendation algorithms and analyzing the association between fairness and network properties over time. We further study how interventions on network properties influence fairness by examining counterfactual scenarios with alternative evolution outcomes and differing network properties. Our results on empirical datasets suggest that recommendation fairness improves over time, regardless of the recommendation method. We also find that two network properties, minority ratio, and homophily ratio, exhibit stable correlations with fairness over time. Our counterfactual study further suggests that an extreme homophily ratio potentially contributes to unfair recommendations even with a balanced minority ratio. Our work provides insights into the evolution of fairness within dynamic networks in social science. We believe that our findings will help system operators and policymakers to better comprehend the implications of temporal changes and interventions targeting fairness in social networks.

Figures (15)

• Figure 1: An illustration of our methodology. We look at a sequence of network snapshots at increasing timestamps. For each snapshot $\mathcal{G}_t$, we train social recommendation models $\mathcal{F}_t$ and obtain the recommended nodes set $R_t$. We address RQ1 (Evolution) by evaluating the recommendation fairness of various algorithms over time on real-world datasets. We address RQ2 (Association) with a correlation study of recommendation fairness and observed real-world network properties. For RQ3 (Intervention), we synthesize alternative sequences of snapshots that simulate counterfactual intervention scenarios, such as increasing the homophily ratio. We train the same recommendation models and evaluate the impact of the intervention on the evolution of fairness.
• Figure 2: Evolution statistics of the empirical datasets.
• Figure 3: Recommendation fairness (VD@100, sVD@100, rVD@100) over time on the empirical datasets. We observe that the recommendation fairness measured by VD and sVD improves over time for all recommendation models, while rVD shows varying trends on the empirical datasets.
• Figure 4: The Spearman correlations between network properties and recommendation fairness (VD) of GCN. We denote the insignificant correlations ($p<0.05$) as dashes (-). We observe that on Norwegian and DBLP, MR, HR, ED-intra, and ED-inter exhibit consistent correlations to recommendation fairness over time, while the results on Enron show no consistent correlations. The results suggest that the evolution of these social network properties may influence recommendation fairness over time.
• Figure 5: Counterfactual network properties. We show the observed minority ratio and homophily ratio on counterfactual networks after intervening on these two properties.

Theorems & Definitions (5)

• Definition 1: Visibility Equality
• Definition 2: Visibility Disparity
• Definition 3: Relative Visibility Disparity
• Definition 4: Maximum Ranking Reciprocal Difference
• Definition 5: Hits Difference