acceptodds
Under review as a conference paper at ICLR 2027

Fair Influence Maximization beyond Predefined Groups

Abstract

Fair influence maximization aims to spread information widely while avoiding large disparities across different parts of a network. Most existing methods assume that protected groups or community labels are known. This assumption is often unrealistic when demographic information is missing, incomplete, sensitive, or defined by multiple overlapping attributes. In addition, raw reach alone may not provide a fair comparison because some nodes are structurally harder to reach than others. We introduce AURA-IM, a framework for fair influence maximization without predefined groups. AURA-IM does not use observed group memberships or the true numbers of attributes/groups during seed selection. Instead, it uses target-specific reverse-reachable samples to estimate how much service each node can realistically attain under a given seed budget. This attainable service is used to define a normalized opportunity signal. A structural multi-view encoder then learns node representations, and a latent auditor adaptively identifies underserved marginal and intersectional groups. Their risk is modeled using KL-robust utility and upper-tail aggregation, and is jointly optimized with influence spread through an adversarial extragradient–Frank-Wolfe procedure followed by discrete refinement. Experiments on eight real-world networks show that AURA-IM achieves strong influence spread while maintaining competitive observed-group fairness. These results suggest that fairness-aware influence maximization is possible even when protected groups are not available during optimization.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Influence Maximization, Latent Groups, Multi-Attribute Fairness, Graph Learning, Adversarial Min-Max Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_wvPAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Fairness in Influence Maximization (IM) tries to spread information widely while avoiding large disparities across different parts of a network. However, existing related methods need to have knowledge about the groups' sensitive attributes, like gender or race, in advance. For this reason, the authors propose Adaptive Unobserved-Group Attainability for Influence Maximization (AURA-IM) for fair influence maximization beyond predefined groups. It normalizing node reach against structural attainability ceilings, it uses an adversarial latent auditor to discover worst-served intersectional latent groups, and uses Frank-Wolfe optimization and extragradient updates to optimize seed selection. The labels of the groups are used only after seed selection, for evaluation.

Critical strengths
  • The paper introduces a meaninguful problem, that of ffair influence maximization without access to nodes’ sensitive attributes, a realistic setting that most existing fair IM methods, which assume known groups, do not cover.
Critical weaknesses
  • The overall presentation of the paper is much lower than the standards of a conference like ICLR.

    • Complex and heavilty presented mathematical formulations, without any explanation and intuition. See for example equations (8) and (11). Moreover, notation is overloaded.
    • Bad quality and squeezed figures (see Figure 2).
  • The method section is very heavily written with little intuition for the key ideas. It describes code details rather than the ideas - is is full of details that matter for running the code but not for understanding the method. Moreover, the evaluation has just one results figure, for a single dataset (even though 9 are presented in Table 1), while the comparison of AURA-IM with baselines lies in the Appendix (Tables 2, 7-10). Reading just the main paper can’t check the authors' claims.

Questions for authors
  • Why are all these datasets presented in Table 1 in the main text, if only one of them is actually used in the main experiments?
  • How were the many hyperparameters of Tables 3 and 4 chosen?
  • How are methods not originally designed for Fair IM, are being instantiated in this problem, like for example CrossWalk?
Reviewer_1ukBOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces AURA-IM, a method for choosing "seed" nodes in a network while seeking to avoid systematically underserving portions of the network—without relying on known demographic, community, or other protected-group labels during optimization. The approach estimates each node’s realistically attainable level of exposure under a fixed seed budget, normalizes actual reach against that baseline, learns structural representations of the network, and uses an adversarial latent-group auditor to identify potentially underserved marginal and intersectional populations. It jointly balances overall diffusion with a robust fairness-risk objective (KL-DRO and CVaR tails), then evaluates outcomes against observed labels only after seed selection. Across eight real-world networks, the authors report that AURA-IM generally retains strong influence spread and achieves competitive post-hoc group-fairness measures.

Critical strengths

1.) Motivated Problem formulation: The authors recognize a limitation of most prior group-aware approaches--namely, that protected-group labels are unavailable, incomplete, or inappropriate to use during seed selection.

2.) Extensive Experimentation: The authors experiment on eight real-world networks, compare against nine baselines at additional diffusion settings, include component ablations, and report multiple fairness measures.

Critical weaknesses

1.) Theoretical support is limited: Proposition 1 is a classical greedy coverage approximation via monotone submodularity and Proposition 2 is a standard Frank–Wolfe linear-oracle result. Could anything be said about the quality of the final discrete seed set, or improvement in observed-group fairness?

2.) Difficult to validate: Because the auditor discovers structural subgroups without access to observed group labels, it is not established that these subgroups recover—or are a reliable proxy for—meaningful real-world disadvantaged groups. The authors appropriately acknowledge that AURA-IM cannot guarantee fairness for arbitrary observed partitions. Controlled evidence relating latent-risk reduction to improvements in held-out group-fairness metrics would strengthen the central claim.

Questions for authors

It's difficult for me to conclude that the method proposed by the authors is a meaningful contribution. Ultimately, the method is a series of many interacting components, with no guarantees provided, so it feels more akin to a heuristic.

Q1) The paper appropriately acknowledges that minimizing latent risk cannot guarantee fairness for every unobserved partition. Within the intended scope, can the authors provide controlled empirical evidence showing when latent-risk minimization predicts improved held-out group fairness—and when it does not? For example, the authors could use semi-synthetic networks with known ground-truth marginal and intersectional groups whose labels are withheld during optimization, and evaluate: (1) whether the learned latent groups recover or meaningfully overlap with hidden disadvantaged groups; (2) whether lower optimized latent risk is associated with improved held-out group-fairness metrics across multiple partitions, homophily levels, and random seeds; and (3) whether AURA-IM outperforms strong label-free structural baselines, with uncertainty intervals.

Reviewer_v9RkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an interesting approach to fair influence maximization without predefined groups. Although the technical formulation appears sound, the intended notion of fairness is insufficiently clarified. In particular, demographic group fairness cannot generally be guaranteed without group information or assumptions linking demographics to graph structure: the same graph can admit different demographic label assignments, for which the same seed set may have very different fairness outcomes. The paper therefore needs stronger justification for why protecting learned structural groups should protect demographic populations.

Critical strengths

S1: The paper addresses a meaningful limitation of existing fair influence maximization methods: demographic group memberships may be unavailable or inappropriate to collect.

S2: Normalizing service by an attainable reference is a thoughtful attempt to measure how easily nodes can be reached.

Critical weaknesses

W1: The paper presents a promising approach to fairness without demographic labels, but the evidence does not yet establish that its learned groups reliably protect groups. Specifically, human-defined groups, such as disability status, income, or ethnicity, may have little correspondence with graph structure, and the same graph can admit different demographic label assignments. Consequently, protecting graph-derived latent groups does not necessarily protect disadvantaged demographic populations.

W2: Although Eq. (14) defines the optimization objectives, the exact notion of fairness studied in this paper remains unclear. In particular, the paper does not adequately justify why protecting learned structural groups should protect demographic populations.

W3: The experiments do not sufficiently validate the central motivation that learned structural groups can substitute for unavailable demographic groups. For example, they neither examine settings with weak structural–demographic alignment nor directly evaluate observed demographic intersections, leaving the claimed demographic and intersectional benefits insufficiently supported.

Questions for authors

Q1: What is the exact notion of fairness studied in this paper? Maximin, disparities, welfare, or something else?

Q2: In practice, demographic groups may be defined by social attributes that have little correspondence with graph structure. How does the proposed method protect such groups, and how do you evaluate its effectiveness when structural–demographic alignment is weak or absent?

Reviewer_prixOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is so obscurely written it is hard to understand anything, including the abstract (e.g. "Their risk is modeled using KL-robust utility and upper-tail aggregation, and is jointly optimized with influence spread through an adversarial extragradient–Frank-Wolfe procedure followed by discrete refinement.").

What I got from it is that they invent a method for fair influence maximization which does not utilize group membership during its computation. They experiment on 8 real-world networks, which all have less than ~6k nodes, which is relatively small for influence maximization. They compare against 9 fair and non-fair baselines, and win in 28/32 cases, including against the non-fair baselines.

Critical strengths
  • fair influence maximization without group membership is an important problem
Critical weaknesses
  • this paper is completely unreadable by a human. The writing uses tons of complex words one after the other, introduce a ton of unneeded notations and their final results are in figure 2, which is barely legible. I couldn't follow what they are doing, despite me having published papers on fair influence maximization and having reread many paragraphs multiple times.
  • an example of this is line 288-302. On line 288 they define mm as a vector, then line 295 mm is redefined as a single number, then on line 301 mm is reredefine as a variant the original vector. Moreover, the writing couldn't be more obscure. "For a nonnegative membership vector m of positive mass, define Pm(v)=mv∑umuP_m(v) = \frac{m_v}{\sum_u m_u}". This could be simplified into "For each attribute A, we define a membership probability vector PAP_A". Then "On a finite positive temperature grid H, the implemented KL-robust lower utility is U(m,o)=max⁡η∈H−ηρ−ηlog(∑v∈VPm(v)e−ov/η)U(m, o) = \max_{\eta \in H} −\eta\rho − \eta log (\sum_{v \in V} P_m(v) e^{−o_v/\eta}) [...] This is a finite-grid evaluation of the dual, not an asserted exact continuous optimization." literally just means they optimize U via grid search on η\eta. "Finite positive temperature grid" is not needed, nor is "This is a finite-grid evaluation of the dual, not an asserted exact continuous optimization." This level of obscure writing is misleading at best.
  • it continues in proposition 2, which is completely trivial, yet for someone who hasn't spent the time to remember their notations and what they're doing might look rigorous
  • last but not least, their fair method beats the non-fair baselines, which show the non-fair baselines are not good enough. I couldn't see an RR baseline for instance, which was hard to find since there is no subsection about the baselines in the main text.

Less important: the list of contributions is not a list of contributions, it's the steps of the method.

Questions for authors
  • Are you sure you followed the conference rules around disclosure of the usage of language models in drafting your manuscript?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.