💻 computer science

Recommendation Fairness in Social Networks Over Time

This paper investigates the temporal evolution of recommendation fairness in dynamic social networks, revealing that fairness generally improves over time and is significantly correlated with network properties like minority and homophily ratios, while demonstrating through counterfactual analysis that extreme homophily can undermine fairness even with balanced group representation.

Original authors: Meng Cao, Hussain Hussain, Sandipan Sikdar, Denis Helic, Markus Strohmaier, Roman Kern

Published 2026-02-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Meng Cao, Hussain Hussain, Sandipan Sikdar, Denis Helic, Markus Strohmaier, Roman Kern

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, ever-growing digital town square called a Social Network. In this town, there's a special job: the Recommendation Mayor. Every day, this Mayor suggests who you should meet, who you should follow, or who you should connect with.

The problem? Sometimes, this Mayor has a bias. Maybe they always suggest the "popular kids" (the majority group) and ignore the "quiet kids" (the minority group). Or maybe they only suggest people who look exactly like you, keeping everyone in their own little bubble.

This paper is like a time-traveling investigation into how fair this Mayor is, not just today, but over many years as the town grows and changes.

Here is the story of their findings, broken down into simple parts:

1. The Mystery: Static Snapshots vs. The Moving Movie

Most previous studies looked at the town square like a frozen photograph. They took one picture of the network, checked if the Mayor was fair, and stopped there.

But real life isn't a photo; it's a movie. People join, leave, make friends, and break up. The authors realized: If we only look at one frame, we miss the whole story. So, they watched the "movie" of three real-world networks (a corporate email network, a board of directors network, and a scientific collaboration network) over time to see how fairness evolved.

2. The Tools: Measuring "Visibility"

To measure fairness, they didn't just ask, "Is everyone happy?" They asked, "Who gets to be seen?"

They used a concept called Visibility Disparity. Imagine a stage with spotlights.

Unfair: The spotlight only shines on the majority group (e.g., men), leaving the minority group (e.g., women) in the dark.
Fair: The spotlight splits evenly, giving everyone a chance to be seen.

They tracked this "spotlight" over time for different types of recommendation algorithms (some smart AI, some simple math).

3. The Big Discovery: Time is on the Side of Fairness

Here is the most surprising plot twist: As the town grew older, the Mayor got fairer.

Regardless of which algorithm they used (the smart AI or the simple math), the recommendations became more balanced over time.

Why? As the network grew, new people joined. These new connections naturally mixed different groups together, breaking down the old, rigid bubbles. It's like adding more water to a cup of coffee; eventually, the coffee gets diluted and more uniform. The natural evolution of the network helped fix the bias.

4. The Two "Knobs" That Control Fairness

The researchers found two specific "knobs" on the town's structure that control how fair the recommendations are:

The Minority Ratio (The Crowd Mix): This is simply the percentage of "minority" people in the town.
- Analogy: If the town is 90% Group A and 10% Group B, the Mayor will naturally suggest Group A more often. But as the town grows and Group B gets bigger (the ratio balances out), the suggestions become fairer.
The Homophily Ratio (The "Birds of a Feather" Factor): This measures how much people stick to their own kind.
- Analogy: Imagine a party where everyone only talks to people who wear the same color shirt. That's high homophily. If people start talking to everyone regardless of shirt color, that's low homophily.
- The Finding: The study found that if the "Birds of a Feather" effect gets too strong (everyone only talks to their own group) OR too weak (everyone is forced to talk to strangers), the recommendations get unfair again. There is a "Goldilocks zone" where mixing is just right for fairness.

5. The "What If" Experiment (Counterfactuals)

To prove their point, the researchers ran a simulation. They asked: "What if we forced the town to change?"

They created alternate realities (counterfactuals) where they artificially forced the town to become extremely segregated (high homophily) or extremely mixed, even if the population numbers stayed the same.

The Result: Even if the town had a perfect 50/50 split of men and women, if the groups refused to talk to each other (extreme homophily), the Mayor became unfair again. The AI would only suggest people from your own group because that's all the data showed.
The Lesson: You can't just fix the numbers (how many people are in each group); you have to fix the connections (who talks to whom).

6. The Takeaway for the Real World

So, what does this mean for us?

Don't Panic About Bias Immediately: If you launch a new social network, it might be biased at first. But as the network grows and people naturally make new friends across different groups, the system often self-corrects and becomes fairer.
Structure Matters More Than Algorithms: You don't always need a super-complex AI to fix fairness. Sometimes, simply encouraging people to connect with those who are different from them (lowering the "Homophily") is enough to make the recommendations fair.
Watch the "Knobs": If you are a platform manager, keep an eye on the Minority Ratio and Homophily Ratio. If the town becomes too segregated, the recommendations will get unfair, no matter how smart your code is.

In a nutshell: This paper tells us that social networks are living, breathing things. While they start with biases, time and natural growth often help smooth things out. However, we must be careful not to let the town split into two isolated islands, or the fairness will vanish again.

1. Problem Statement

Social recommender systems often exhibit biases against specific demographic groups (e.g., gender, race), leading to issues like the "glass ceiling" effect (where minorities receive fewer recommendations) and the "rich get richer" phenomenon. While existing research has extensively studied fairness in static social networks, there is a significant gap in understanding how recommendation fairness evolves over time in dynamic networks. Real-world social networks are not static; they undergo continuous changes in nodes, edges, and structural properties. The authors aim to bridge this gap by investigating:

How recommendation fairness evolves over time across different models.
The association between dynamic network properties and fairness.
How structural interventions (counterfactual scenarios) impact fairness evolution.

2. Methodology

Datasets and Setup

The study utilizes three real-world dynamic social networks with diverse growth patterns:

Enron: Email communication network (28 months).
Norwegian: Board of directors network (112 months).
DBLP: Author collaboration network (18 years).
Sensitive Attribute: Gender (inferred from names or original data), partitioning nodes into Majority ( $M$ ) and Minority ( $m$ ) groups.

Fairness Metrics

The authors define fairness primarily through Visibility Disparity (VD), measuring the discrepancy in recommendation probability between groups.

Visibility Equality: $P(group(r)=M) = P(group(r)=m)$.
Visibility Disparity (VD): The absolute difference between these probabilities. Lower VD indicates higher fairness.
Signed VD (sVD): Indicates which group is favored (positive = majority favored).
Relative Visibility Disparity (rVD): Adjusts for group size imbalances in the population.
Extended Metrics: The paper also analyzes ranking fairness (MRRD) and utility fairness (HitsDiff) in appendices.

Recommendation Models

Six algorithms were evaluated across three categories:

Baseline: Random recommendation.
Graph Neural Networks (GNNs): GCN, LightGCN, and FairDrop (a fairness-aware model).
Matrix Factorization / Collaborative Filtering (MF/CF): BPR, NeuMF, NGCF.

Experimental Approach

The study addresses three Research Questions (RQs):

RQ1 (Evolution): Models are trained on network snapshots at time $t$ and evaluated on links at $t+1$ . Fairness metrics are tracked over the entire timeline.
RQ2 (Association): Spearman and Pearson correlation analyses are performed between fairness metrics and 10 network properties (e.g., Minority Ratio, Homophily Ratio, Edge Density, Clustering Coefficient).
RQ3 (Intervention): A Bernoulli-Barabasi-Albert (BBA) model is used to generate counterfactual networks. The authors synthetically manipulate the Minority Ratio (MR) and Homophily Ratio (HR) to extreme values (0.1 to 0.9) while keeping other properties neutral, observing how these structural changes alter fairness evolution.

3. Key Contributions

Temporal Analysis of Fairness: The paper is one of the first to systematically study the evolution of recommendation fairness in dynamic networks rather than static snapshots.
Identification of Key Drivers: It identifies Minority Ratio (MR) and Homophily Ratio (HR) as the two most critical network properties correlated with fairness, outperforming other structural metrics like edge density or clustering coefficients in predictive power.
Counterfactual Intervention Framework: The authors propose a novel methodology using the BBA model to simulate "what-if" scenarios, isolating the causal impact of specific structural changes (MR and HR) on fairness.
Comprehensive Evaluation: The study evaluates a wide range of models (from simple random to advanced GNNs) across multiple datasets and fairness definitions (absolute, signed, and relative).

4. Key Results

RQ1: Fairness Evolution

Improvement Over Time: Contrary to the assumption that bias persists or worsens, recommendation fairness (measured by VD) improves over time for all recommendation methods across all three datasets.
Model Performance: Most advanced models (GNNs, MF/CF) achieve lower disparity than random recommenders, though the trend of improvement is consistent regardless of the algorithm used.
Relative Fairness: When accounting for group sizes (rVD), the disparity is lower, and the trend is less pronounced, suggesting that algorithms naturally adapt to the changing minority ratio.

RQ2: Association with Network Properties

Strong Correlations: There is a stable, strong correlation between fairness and Minority Ratio (MR) and Homophily Ratio (HR).
- MR: Higher minority ratio correlates with better fairness (lower VD).
- HR: Lower homophily (more heterophily) correlates with better fairness.
Predictive Power: Regression analysis (Random Forest) shows that MR and HR alone are sufficient to accurately predict fairness, outperforming models that include all other network properties.

RQ3: Impact of Interventions (Counterfactuals)

Minority Ratio (MR): Changing MR linearly results in a linear change in fairness. Higher MR consistently leads to fairer outcomes.
Homophily Ratio (HR): The impact of HR is non-linear and critical.
- Extreme Values are Detrimental: Both extremely high homophily (segregation) and extremely low homophily (bipartite-like structures) lead to worse fairness, even if the group sizes (MR) are perfectly balanced (0.5).
- Mechanism: Extreme homophily creates isolated components where only majority nodes are recommended to majority nodes. Extreme heterophily creates a bipartite structure where recommendations are strictly cross-group, potentially starving one group of internal connections.
- Optimal Zone: Moderate levels of homophily yield the best fairness.

5. Significance and Implications

For System Operators: Simply deploying fairness-aware algorithms is not enough; the underlying data dynamics matter. Operators should monitor the evolution of MR and HR. If a network becomes too segregated (high homophily), fairness will degrade even if the group sizes are balanced.
Policy and Intervention: The study suggests that structural interventions can improve fairness. For example, in highly homophilic networks, encouraging cross-group connections (reducing homophily) can improve fairness. Conversely, in highly heterophilic networks, fostering some intra-group connections might be beneficial.
Long-term Perspective: The findings challenge the static view of fairness. They suggest that as social networks naturally evolve (often becoming more balanced and less homophilic over time in the studied datasets), fairness may improve organically. However, artificial interventions must be carefully calibrated to avoid extreme structural states.

In conclusion, the paper establishes that recommendation fairness is a dynamic property deeply intertwined with the temporal evolution of network structure. It provides a framework for understanding and intervening in these dynamics to ensure equitable visibility for all demographic groups in social networks.