acceptodds
Under review as a conference paper at ICLR 2027

Attribute-Balanced Fair Influence Maximization via Strategic Group Competition

Abstract

Fair influence maximization aims to select influential seed nodes while ensuring that information spreads fairly across different groups. Existing methods often define fairness over a single group partition or treat all groups as a flat collection. This becomes limiting when several categorical attributes are considered at the same time, because attributes with more groups may have greater influence on the selection process, and raw group utilities may not distinguish between persistent under-service and limited structural opportunity. We propose SAGE-IM, a multi-attribute fair influence maximization framework in which each attribute-group pair is modeled as a strategic agent. SAGE-IM combines opportunity-normalized deficits, cross-step regret, and hierarchical attribute balancing to adapt fairness pressure as the seed set grows. Group-specific teachers learn RR-based marginal utilities, and their preferences are distilled into a shared student through policy, ranking, and representation objectives. During sequential selection, the learned models are used to generate a compact candidate set, while the final seed choice is based on current global and group-specific reverse-reachable marginals. This design reduces reliance on learned rankings and keeps the final decision tied to the sampled diffusion structure. Experiments on eight processed real-world networks show that SAGE-IM achieves strong attribute-wise fairness and worst-group coverage while maintaining competitive influence spread. Sensitivity and ablation analyses further show how fairness dynamics, strategic coordination, attribute balancing, and active-agent selection shape the spread-fairness trade-off.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Influence Maximization, Multi-Attribute Fairness, Multi-Agent Competition, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mofROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper address the problem of fair influence maximisation with fairness to be accounted for across multiple attributes. It models each attribute-group pair as a strategic agent. Group-specific “teacher” networks learn RR-based marginal utilities and are distilled into a shared student. The learned models only propose candidates. The final seed choice uses exact sampled reverse-reachable (RR) marginals, so a model misranking can’t by itself exclude strong nodes. On eight real networks, SAGE-IM generally achieves higher entropy and worst-group coverage with competitive spread, though not uniformly compared with existing IM algorithms.

Critical strengths

-I appreciate the novelty of studying of IM with fairness across different attributes -I appreciate the honesty of the authors in what the paper claims to cover, highlighting the limitations: the RR expectation gives no uniform-accuracy guarantee under capped banks, the weight identities are not fairness guarantees, and the gradient projection only attenuates conflict rather than guaranteeing better spread. The problem is notoriously hard to solve.

Critical weaknesses

The method has many complicated parts, but the paper’s own tests don’t show that they’re all needed. Switching some parts off barely changes the results, and sometimes even improves them.

Questions for authors

Q: how does SAGE-IM compare against a simple baseline that skips the learned teacher and student models and picks seeds directly from the group and global RR marginals with the same fairness weighting?

Reviewer_TTjTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies fair influence maximization with multiple categorical attributes. SAGE-IM coordinates attribute–group agents through coverage deficits, regret updates, and hierarchical weighting. Group-specific teachers and a distilled student generate candidates, while the final seed selection uses current marginal gains from sampled RR sets. Experiments on eight processed networks show improvements in several spread and fairness measures.

Critical strengths

The motivation is reasonable: pooling all groups together can give disproportionate weight to attributes with more categories. Using RR marginals for the final decision is also a sensible way to reduce reliance on learned predictions. I appreciate that the evaluation considers entropy, worst-group coverage, and dispersion together, and acknowledges that these measures do not always agree.

Critical weaknesses
  1. My main concern is the experimental comparison. The paper includes several baselines, but lacks comparisons with more recent fair influence maximization methods, such as FIMM (Scalable Fair Influence Maximization, NeurIPS 2023) and other similar methods. FIMM is particularly relevant because it also uses reverse influence sampling to optimize fairness and influence spread. Without such comparisons, it is difficult to judge the benefit of the proposed coordination mechanism over existing RR-based fair IM methods.
  2. The datasets are also small for assessing scalability. All evaluated networks contain at most 6,040 nodes. The JAIR 2023 paper (Fair influence maximization in large-scale social networks based on attribute-aware reverse influence sampling) on attribute-aware reverse influence sampling already provides experimental benchmarks with hundreds of thousands to over a million nodes. Its large-scale benchmarks include overlapping groups induced by multiple attributes. Thus, the multi-attribute setting alone does not explain the limited experimental scale here. Results on larger attributed networks, together with runtime and memory measurements, would substantially strengthen the evaluation.
  3. The theoretical analysis mainly establishes basic RR properties and algebraic identities for the weighting and gradient updates. It does not establish guarantees for the final selection procedure, which combines capped sampling, candidate truncation, dynamic weights, and modal-action selection. In particular, the regret updates are used as a heuristic coordination signal, but their connection to fairness or convergence is not established. The description as “strategic competition” would benefit from a clearer game formulation or an analysis of the resulting procedure.
  4. The contribution of the teacher–student architecture is not sufficiently demonstrated. RR marginals are already maintained and used for candidate generation and final scoring, while some ablation curves without the learned proposal models are close to the full method. A comparison under matched RR and candidate budgets, reporting both performance and computational cost, is needed to justify this added complexity. The reduction in active agents alone does not demonstrate an overall speedup, since all teachers still score the remaining nodes and all group trackers are updated.
Questions for authors
  1. How does SAGE-IM compare with recent fair IM methods such as FIMM?
  2. How does the method perform on larger networks with overlapping attribute groups, and what are its runtime and memory costs?
  3. Under matched sampling and candidate budgets, how much do the learned proposals improve over RR-only proposals?
  4. Can the authors clarify what the regret-based procedure guarantees, and whether “strategic competition” refers to a formally defined game or a heuristic weighting mechanism?
Reviewer_6MAqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAGE-IM, a multi-attribute fair influence maximization framework based on strategic competition among attribute-group agents. The method integrates opportunity-normalized fairness deficits, regret-based competition, hierarchical attribute balancing, and teacher-student distillation. Reverse-reachable sampling is employed for candidate evaluation and seed selection. Experiments on eight networks demonstrate improvements in fairness-related metrics while maintaining competitive influence spread.

Critical strengths
  1. The paper addresses multi-attribute fairness, an important extension of conventional single-attribute fair influence maximization.
  2. The combination of learned candidate generation and RR-based marginal evaluation is reasonable and helps maintain reliable seed selection.
  3. The paper considers multiple fairness metrics and provides ablation and sensitivity analyses to investigate the contributions of different components.
Critical weaknesses
  1. Insufficient methodological novelty and theoretical analysis. The distinction between SAGE-IM and existing fairness-aware IM methods is not sufficiently clear. Moreover, the paper lacks a rigorous analysis of the proposed framework's approximation properties and computational complexity.
  2. Insufficient comparison with recent methods. The latest experimental baselines are from 2024, although several relevant studies published in 2025 and 2026 are discussed in the references. Comparisons with these recent approaches would strengthen the evaluation.
  3. Limited experimental scalability and efficiency evaluation. The eight datasets contain only 327-6,040 nodes, and the selection criteria are insufficiently justified. Furthermore, systematic runtime and memory comparisons with competing methods are missing.
  4. Presentation and experimental organization need improvement. Figure 1 is looks blurry, and the pseudocode formatting should be improved. The main paper contains relatively limited experimental results, with substantial supporting analyses placed in the appendix.
Questions for authors
  1. What are the fundamental methodological differences between SAGE-IM and recent fairness-aware IM approaches, particularly those published in 2025 and 2026?
  2. Can the authors provide theoretical guarantees or complexity analyses for the proposed framework?
  3. Why were these relatively small datasets selected? Can the authors provide experiments on larger networks and comparisons of runtime and memory consumption?
  4. Can the authors include more recent baselines and provide a simplified version of SAGE-IM to better justify its architectural complexity?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.