Source-linked AI summary
Algorithmic Fairness for Influence Maximization: A Survey of Models, Metrics, and Optimization Techniques
Shashank Sheshar Singh, Sunil Kumar Meena, Kuldeep Singh, krati Saxena
TL;DR
Classical influence maximization can overlook fairness and leave marginalized or underrepresented communities disadvantaged. This survey organizes fairness-aware models, metrics, and optimization methods and examines their evaluation. Its empirical analysis highlights parity seeding and diversity constraints in multi-objective fairness settings.
Problem
Classical IM prioritizes overall influence spread and can overlook marginalized or less-connected communities and fair distribution across groups.
Source: §1 · pp. 1–2Method
The survey reviews fairness-aware IM models, metrics, optimization methods, and bias-mitigation approaches, and categorizes methods by fairness objectives and algorithmic strategies.
Source: §1.1 · p. 2Results
The survey’s empirical analysis highlights parity seeding and diversity constraints in multi-objective fairness settings.
Source: §8 · p. 30Takeaways & Limitations
The survey emphasizes balancing influence performance with equitable impact in contexts involving social, economic, or demographic disparities.
Source: §8 · p. 30Reliable protected-attribute labels are scarce in widely used real-world network benchmarks, limiting reproducibility and leading some studies to assume attributes or rely on small domain-specific datasets.
Source: §6.4 · p. 25
Abstract
from arXiv · showhide
Influence maximization (IM) has become a basis of modeling information spread across social networks, with applications in viral marketing, public health, and political campaigns. However, classical IM algorithms often prioritize the spread of overall influence and ignore its fairness, resulting in disadvantages for marginalized or underrepresented communities. This survey provides a detailed look at fairness-aware influence maximization, focusing on important challenges, models, and methods designed to ensure fair influence maximization. We categorize existing methods into five groups based on their fairness objectives, including parity, diversity, and utility, and discuss how they handle biased network structure, protected attributes, and seeding constraints. We also examine how fairness-aware influence maximization methods perform on real-world data to highlight the importance of combining fairness with classical influence maximization methods. By bridging the gap between influence spread and social equity, this survey lays the foundation for designing fair algorithms that ensure fair representation in large-scale diffusion processes. The survey concludes by outlining emerging trends, open challenges, and prospective future directions in the domain, highlighting the need for expertise spanning multiple disciplines.
1 INTRODUCTION
Classical influence maximization prioritizes reach and often overlooks marginalized or less-connected communities, protected attributes, and fair distribution. This survey reviews fairness-aware approaches, their objectives and methods, and how they differ from earlier surveys of influence maximization.
1 INTRODUCTION
- Classical IM favors highly connected or visible nodes and can overlook minority and less-connected communities, protected attributes, and fair distribution of influence.Its assumptions about influence probabilities and limited seed constraints also leave important social and cultural variation unmodeled.Source: §1 · p. 2
- Fairness-aware IM considers inclusiveness and representational balance alongside reach, with fairness requirements shaped by application context.The passage contrasts equal community coverage in public health with equal seed counts across groups on social media.Source: §1 · p. 2
1.1 Key Contributions
- The survey presents a taxonomy of fairness-aware methods and compares how they address seed-set, influence-spread, and individual fairness.It also discusses fairness-aware attributes as a way to extend classical IM beyond maximizing reach.Source: §1.1 · p. 2
- The survey examines datasets, evaluation, and reproducibility, and outlines emerging trends, open challenges, and future directions.Source: §1.1 · p. 3
1.2 Difference from Existing Surveys
- Earlier surveys generally address algorithmic, computational, or application-specific IM without examining social diffusion fairness; two other surveys treat fairness more narrowly.This survey focuses on fairness metrics, equitable seed selection, and bias mitigation.Source: §1.2 · p. 3
2 PRELIMINARIES
Classical IM selects a fixed-size seed set to maximize expected influence under a diffusion model. Its established algorithms balance spread and computational cost, while fairness-aware IM adds equitable distribution across social groups.
2.1 Influence Maximization (IM)
- IM selects a seed set of size k to maximize expected influence spread; the problem is NP-hard, so solutions commonly seek near-optimal sets with approximation, heuristic, or learning-based methods.The expected spread is defined under diffusion models such as independent cascade and linear threshold.Source: §2.1 · p. 4
2.2.1 Independent Cascade Model.
- In the independent cascade model, each newly activated node gets one probabilistic chance to activate neighbors, and the process ends when no further activations occur.Source: §2.2.1 · p. 4
2.2.2 Linear Threshold Model.
- In the linear threshold model, a node activates when the proportion of active neighbors exceeds its activation threshold.Source: §2.2.2 · p. 4
2.3 Classical IM Algorithms
- Classical IM algorithms include greedy, heuristic, sampling, and metaheuristic approaches, trading influence estimation quality against computational cost.Greedy methods can be time-consuming, while local-feature heuristics and sampling methods aim to reduce computation; these approaches generally assume neutral conditions and overlook fairness, trust, and diversity.Source: §2.3 · pp. 4–5
2.4 Defining Fairness-aware Influence Maximization
- Fairness-aware IM extends spread maximization by seeking fair influence distribution across groups, including demographic, location-based, and behavioral communities.Source: §2.4 · p. 5
3 TAXONOMY OF FAIRNESS-AWARE INFLUENCE MAXIMIZATION
The taxonomy distinguishes fairness by protected groups, individuals, causal mechanisms, and population-wide distributions, then describes optimization objectives and operational constraints. It also contrasts preprocessing, in-processing, and post-processing approaches, with most existing methods falling under in-processing.
3.1.1 Protected Attributes and Groups.
- Protected attributes such as race, gender, age, and socioeconomic status may correlate with network position, leaving marginalized groups with fewer connections and less influence.Group fairness seeks equitable influence opportunities across protected groups.Source: §3.1.1 · pp. 5–6, §3.1.2 · p. 6
3.1.3 Individual Fairness.
- Individual fairness asks whether people with similar features or network positions receive similar chances of being influenced.Source: §3.1.3 · p. 6
3.1.4 Causality-based Fairness.
- Causal fairness examines whether protected attributes lead to disparities in influence, distinguishing those effects from false correlations.Source: §3.1.4 · p. 6
3.1.5 Distributional Fairness.
- Distributional fairness evaluates the overall allocation of influence, using measures such as max-min fairness, proportionality, and the Gini index.Source: §3.1.5 · p. 6
3.2 Fairness-aware IM Objectives
- Fairness-aware objectives include hard constraints, Pareto trade-offs, unfairness penalties, max-min optimization, and composite criteria.Hard constraints may reduce spread; regularized objectives depend on tuning λ, while max-min objectives prioritize the worst-off group and may limit spread.Source: §3.2 · pp. 6–7
3.3.1 Seed Selection-based.
- Seed-selection measures include demographic parity and cost-aware fairness, while outreach measures compare per-capita group influence, maximize the least-served group, or align influence with group size.Estimating group influence under fairness constraints can make these outreach metrics difficult to optimize.Source: §3.3.1 · p. 7, §3.3.2 · p. 7
3.3.3 Individual Characteristics-based.
- Individual-characteristic measures include influence equity for nodes with similar influence and metric-based fairness for nodes close in a fairness-aware metric space.The section identifies defining similarity in a networked context as a challenge.Source: §3.3.3 · p. 7
3.4 Algorithmic Approach
- Pre-processing modifies data, post-processing adjusts outputs, and in-processing incorporates fairness into learning or optimization; most existing fairness-aware IM methods are in-processing.These methods commonly use constraints, multi-objective optimization, heuristics, regularization, or approximation.Source: §3.4 · p. 8
4 ALGORITHMIC CLASSIFICATION OF FAIRNESS-AWARE IM METHODS
The section reviews fairness-aware IM literature using a six-class categorization and summarizes methods, fairness functions, and evaluation dimensions.
4 ALGORITHMIC CLASSIFICATION OF FAIRNESS-AWARE IM METHODS
- The review categorizes existing fairness-aware IM literature into six classes.Source: §4 · p. 8
- Table 3 compares algorithmic classification, evaluation criteria, datasets, baselines, research challenges, future directions, and fairness considerations.Source: §4 · p. 8
- Table 2 summarizes fairness functions, their parameters, and evaluation metrics used in existing literature.Source: §4 · p. 8
4.1 Constrained Optimization Methods
Constrained methods encode fairness through limits on group representation or influence while maximizing spread. The reviewed work spans demographic-aware seeding, group-fair optimization, and approximation algorithms.
4.1 Constrained Optimization Methods
- Constrained optimization limits group representation or influence during seed selection, using greedy algorithms or integer programming while facing scalability challenges on large networks.The methods aim to maximize spread subject to fairness requirements.Source: §4.1 · p. 8
4.1.1 Demographic / Parity-Constrained IM.
- Feature-aware seeding in can improve fairness and efficiency; in, diversity seeding outperformed agnostic seeding with 9,118 seeds versus 11,286 for parity seeding.The studies examine demographic inequalities and biased networks, respectively.Source: §4.1.1 · p. 8
4.1.2 Maximin / Group-Fairnes Constrained IM.
- A multi-objective submodular framework reduces fairness violations by standard methods, at a small cost to overall utility.It formalizes maximin fairness for the least-advantaged group and diversity constraints tied to group size and internal connectivity.Source: §4.1.2 · p. 9
4.1.3 LP / Approximation-based Constrained IM.
- [18]’s randomized FairCS algorithm reduced unfairness to less than 0.05% while significantly increasing content spread on real datasets.The problem selects candidate edges for fair content spread and is NP-hard and inapproximable in polynomial time unless P = NP.Source: §4.1.3 · p. 9
4.2 Multi-Objective Optimization Methods
Multi-objective methods treat fairness and influence spread as competing goals, seeking balanced solutions without rigid constraints. The reviewed work includes Pareto-style trade-offs, a multi-objective framework, and welfare-based approaches.
4.2.1 Pareto-based Fairness–Influence Trade-off.
- Balanced seed sets can enhance overall influence, especially under high homophily, while Disparity Seeding targets gender ratios and information spread.[12] uses gender-aware metrics and simulation-based seed selection to meet target ratios while maximizing spread.Source: §4.2.1 · p. 9
- MaxFair measures unequal information access, while fairness-aware blocking in uses maximin protection across communities and is evaluated on six real-world datasets.[28] introduces FIBMM, a sampling-based approximate algorithm for the NP-hard fairness-aware influence blocking problem.Source: §4.2.1 · pp. 9–10
4.2.2 Metaheuristic Multi-objective.
- FIMMOGA uses prior knowledge to generate diverse solutions balancing influence and fairness, with the Gini coefficient among its fairness metrics.The problem is described as NP-hard.Source: §4.2.2 · p. 10
4.2.3 Welfare-based Optimization.
- The welfare approach in models utility from information access rather than relying solely on sensitive attributes, and invokes the difference principle to benefit the least advantaged.The cited discussion also connects network position to advantages and disadvantages through the strength of weak ties.Source: §4.2.3 · p. 10
- Fairness trade-offs can be controlled by inequality aversion, with Price of Fairness measuring efficiency loss; SetMOGWO optimizes influence and fairness metrics including propagation speed.The cited discussion says fairness is domain-dependent and requires attention to community structures and decision-making processes.Source: §4.2.3 · p. 11
4.3 Regularization Methods
Regularization places fairness penalties in the objective or modifies diffusion dynamics, with outcomes sensitive to hyperparameter selection. The reviewed formulations include demographic parity and relaxed fairness constraints.
4.3 Regularization Methods
- Regularization penalizes unfairness in the objective, while in-processing can modify diffusion dynamics through biased edge weights; results depend on hyperparameters and may lack interpretability.The fairness penalty can target disparities in group influence.Source: §4.3 · p. 11
4.3.1 Demographic Parity Regularization.
- Strict demographic parity can cause substantial losses in overall spread and algorithm runtime, and its Price of Fairness may be unbounded.The reviewed algorithms target different strictness levels; experiments report the best demographic-parity values with spread comparable to efficient unconstrained algorithms.Source: §4.3.1 · p. 11
4.3.2 Surrogate / Relaxed Fairness Optimization.
- FAIR-TCIM-BUDGET and FAIR-TCIM-COVER reduce group disparities, although enforcing fairness may slightly reduce overall performance.The methods use surrogate objectives with submodular properties for budget- and coverage-constrained time-critical influence maximization.Source: §4.3.2 · p. 11
- [99] extends the IC model with deadlines and random meeting events, and evaluates seed-set quality by expected influence spread using Monte Carlo simulations.The cited account presents MIIA as outperforming standard heuristics for influence maximization.Source: §4.3.2 · p. 11
4.4 Pre-processing and Post-processing Methods
Pre-processing modifies network data before seed selection, whereas post-processing adjusts seeds after optimization; the reviewed diversity-oriented methods also incorporate community and budget considerations.
4.4 Pre-processing and Post-processing Methods
- Pre-processing and post-hoc adjustments are computationally efficient but may not address core causes of bias in influence diffusion.Pre-processing debiases network data before seed selection; post-hoc methods adjust seeds after optimization to meet fairness criteria.Source: §4.4 · p. 12
4.4.1 Diversity Rebalancing.
- DBIM integrates diversity and budget constraints, outperforming existing methods on activated nodes and community influence across all eight datasets.Its objective maximizes diversity and active nodes; the method also reduces overlap among active nodes when searching for seeds.Source: §4.4.1 · pp. 12–13
- DCDIM dynamically detects communities and selects seeds to balance influence spread with diversity among influenced nodes.Its objective is described as monotone and submodular under the linear threshold model.Source: §4.4.1 · p. 13
- DCDIMB uses bridge nodes to diversify active communities and reports that its objective maximizes the number of communities compared with benchmark algorithms.The work introduces a diffusion model incorporating inactive-node influence and proves monotonicity and submodularity under that model.Source: §4.4.1 · p. 13
4.5 Scalable Approximation Techniques
This section examines how fairness-aware influence maximization can scale to large networks while balancing group fairness with total influence spread. It covers sampling, heuristic pruning, community structure, and poverty reward’s concave fairness framework.
4.5 Scalable Approximation Techniques
- Sampling, heuristic pruning, and community-aware approximations reduce complexity for large networks but trade theoretical guarantees for real-world practicality.Fair-RIS is given as an example of sampling-based methods.Source: §4.5 · p. 13
- Poverty reward balances fairness with total influence spread, while its concave fairness framework allows any concave function to mitigate group-level unfairness.The framework quantitatively relates concave-function characteristics to fairness levels and contrasts with approaches focused on a single fairness aspect.Source: §4.5 · p. 13
- With all three seed choices increasing total utility by 0.2, poverty reward favors the fairer utility vector <0.5, 0.4, 0.2> over the alternatives.The alternatives are <0.6, 0.4, 0.1> and <0.6, 0.4, 0.1>; classical seed selection treats all three choices equally.Source: §4.5 · p. 13
- Challenges include improving fairness metrics and scalability in dynamic networks, as well as the NP-hard task of jointly maximizing influence and fairness.Source: §4.5 · p. 15
- Fairness-aware influence maximization faces a trade-off between maximizing influence and ensuring fairness across communities.The survey identifies balancing fairness with total influence spread as a challenge.Source: §4.5 · p. 16
- Scalable fairness-aware embedding methods for large dynamic networks are identified as a direction for future work.Source: §4.5 · p. 16
- [24] is described as the first policy-driven solution applicable at scale to maximize influence while ensuring fairness across groups.Source: §4.5 · p. 16
- Existing methods may fail on large datasets, and scalability, NP-hardness, and robustness guarantees for sampling remain challenges.Some methods were tested only on small networks.Source: §4.5 · p. 16
- Future work includes more effective sampling approaches and new diffusion models to capture network dynamics.Source: §4.5 · p. 16
4.6 Network Embedding and Learning-based Approaches
Network embeddings and learning-based methods extend influence maximization with fairness-aware representations, seed selection, or rewards. The section also covers learning-based approaches to influence blocking.
4.6.1 Network Embedding-based Approaches.
- CrossWalk biases random walks across group boundaries and selects separated seed nodes, while MoonWalk refines boundary proximity for networks with multiple groups.Both aim to promote fairer influence maximization while maintaining influence spread performance.Source: §4.6.1 · p. 18
- Fairness-based participant sampling learns embeddings and diffusion probabilities independently, whereas fairness as context jointly learns fairness and information cascades.A fair-greedy algorithm selects seeds using embeddings and diffusion probabilities; coefficient of variation measures fairness, with equal influenced proportions indicating perfect fairness.Source: §4.6.1 · p. 18
4.6.2 Learning-based Approaches.
- DQ4FairIM adds max-min fairness reward components to reinforcement learning, balancing overall spread with equitable outreach through reward shaping.Other described influence-maximization methods combine Structure2Vec with deep Q-learning or use reinforcement learning to estimate expected influence and generalize across similar network types.Source: §4.6.2 · p. 18
- For influence blocking, reinforcement learning models intervention as sequential node selection, with rewards defined by reductions in expected adversarial spread.Value-based, policy-gradient, and actor–critic methods learn adaptive strategies that can generalize across graph structures and adversarial scenarios.Source: §4.6.2 · p. 19
- Adversarial embeddings aim to reduce disparity while maintaining competitive influence spread; a centrality-based embedding method selects truth seeds without diffusion models.The latter was evaluated on four real-world networks against five baselines and outperformed them in blocking while reducing community fairness discrepancies.Source: §4.6.2 · p. 19
- A reinforcement-learning fairness method uses a generalized Gini welfare function to balance fairness and utility, with experiments reported as validation.The section also describes the Pigou–Dalton principle: transferring rewards from a better-off group to a worse-off group is considered fairer.Source: §4.6.2 · p. 19
5 DATASETS AND EVALUATION METRICS
The datasets discussion identifies limited access to protected attributes and the resulting reliance on simulated or inferred group information. The section also surveys evaluation metrics and identifies the metric used in the survey’s experiments.
5.1 Datasets
- Commonly used datasets often lack protected attributes, so researchers simulate or infer group information, which may introduce noise or bias.The section notes a need for ethically sourced datasets with broad user attributes, ground-truth group labels, and true impact propagation traces.Source: §5.1 · pp. 19–20
- Table 4 summarizes dataset information, including protected attributes.The survey also identifies a need for datasets with ethical sourcing, ground-truth groups, and propagation traces.Source: Table 4 · p. 20, §5.1 · p. 20
5.2 Evaluation Metrics
- Table 5 surveys fairness evaluation metrics for influence maximization, including their formulas, descriptions, advantages, and disadvantages.The survey reports using the most widely used metric in its experiments, without naming it in these passages.Source: §5.2 · p. 20, Table 5 · p. 20
6 EMPIRICAL EVALUATION AND REPRODUCIBILITY
The evaluation compares fairness-aware and classical influence-maximization algorithms across five social-network datasets using influence-spread and fairness metrics. Fairness-aware methods often preserve influence spread while improving community coverage or equity, but results vary by metric and network; reproducibility remains challenging.
6.1 Experimental Set-Up
- Experiments configure graph attributes, including influence weights and Louvain communities, run benchmark algorithms to obtain seed sets, then assess spread and fairness metrics.The compared set includes fairness-based and non-fairness-based algorithms; evaluation uses five datasets and the ICM model with influence probability 0.85 and seed sizes {2, 4, 8, 16, 32}.Source: §6.1 · p. 21, §6.2 · p. 21, §6.3 · p. 21
6.3.1 Influence spread.
- Across datasets, fairness-based algorithms achieve influence spread equivalent to other algorithms, while greedy-based methods remain competitive and PageRank has the lowest spread.Greedy and CELF perform especially well on Facebook and Twitter; spread rises as seed-set size increases, though its scale varies by network.Source: §6.3.1 · p. 22, Table 6 · p. 22
6.3.2 Diversity (number of communities).
- PDTF maximizes community coverage on Twitter, while DBIM reaches up to 30.0 communities and DGA up to 41.0 on LastFM; Hub and Authority show low diversity.Greedy approaches perform well in lower-complexity settings but fail to capture community breadth on LastFM.Source: §6.3.2 · pp. 22–23, Table 7 · p. 22
6.3.3 Equality.
- Equality scores are close to zero overall and tend to decline as seed sets grow from k=2 to 32, although diversity-based methods perform slightly better.Bridge and DCDIM show a sharp increase at k=32 on Twitter; DBIM and CDIM are more sensitive to graph structure.Source: Table 8 · p. 23, §6.3.3 · p. 24
6.3.4 Equity.
- Fairness-aware algorithms show the highest equity values, while centrality-based algorithms have very low equity as seed sets scale.DCDIM and CDIM reach relatively high equity on Facebook, whereas Congress has lower equity values.Source: Table 9 · p. 23, §6.3.4 · p. 24
6.3.5 Disparity.
- PDTF maintains a constant, significantly lower disparity penalty on Twitter, while PageRank and Authority exceed 2250.0; lower penalties indicate more equitable distribution.DCDIM and CDIM reduce penalties as seed sets grow on dense Facebook; Hub and Authority show high penalties on LastFM.Source: Table 10 · p. 24, §6.3.5 · p. 25
6.3.6 Key empirical observation on fairness-based algorithms.
- DCDIM and CDIM match or nearly match Greedy's influence spread, while DCDIM, DBIM, and Bridge reach more communities and DCDIM, CDIM, and DBIM yield lower disparity than Greedy and CELF.SSR-PEA and FIMMOGA perform worse in spread; PDTF often has high, constant diversity but flat spread, possibly reflecting a sacrifice in spread from selecting seeds from maximum communities.Source: §6.3.6 · p. 25
6.4 Reproducibility Challenges
- Reproducibility is constrained by scarce reliable protected attributes in real-world datasets and sensitivity of fairness metrics to simulation and diffusion settings.The paper calls for shared datasets, standardized benchmarks and metrics, and transparent reporting of experimental settings.Source: §6.4 · p. 25
7 EMERGING TRENDS, OPEN CHALLENGES, AND FUTURE DIRECTIONS
The survey identifies fairness challenges that span changing networks, causal and contextual definitions, uncertainty, scalability, and real-world impact. It outlines future approaches that combine equity with practical constraints, stronger evaluation, complex diffusion models, and interdisciplinary input.
7.1.1 Dynamic/Adaptive Fair IM; 7.1.2 Causal Fairness in IM
- Adaptive policies can revise seed selection as networks change, while causal fairness uses path-specific effects to distinguish legitimate network roles from unfair influence through protected attributes.Structural causal models are proposed for identifying and mitigating unfair dependencies.Source: §7.1.1 · pp. 25–26, §7.1.2 · p. 26
7.1.3 Multi-sided Fairness; 7.1.4 Combining Fairness with Other Objectives
- Multi-sided fairness considers receivers, providers, and sources, while combining fairness with robustness, cost, and reach creates trade-offs that motivate Pareto-oriented optimization.Maximizing receiver diversity can reduce provider utility, and standalone fairness constraints may reduce total influence.Source: §7.1.3 · p. 26, §7.1.4 · p. 26
7.1.5 Defining Context-Aware Fairness; 7.1.6 Mitigating Underlying Network Bias
- Context-aware fairness varies priorities by information and socio-technical setting, while structural reweighting or fair graph embeddings aim to bridge network divides at possible cost to total spread.Context-dependent utility functions and formal mappings from social norms to constraints remain challenges.Source: §7.1.5 · p. 27, §7.1.6 · p. 27
7.1.7 Handling Uncertainty in Group Membership and Network Structure
- Uncertain group labels, graph structure, and influence probabilities motivate robust seed selection across possible scenarios rather than optimization for one assumed network.Distributionally robust optimization and interval-based reverse influence sampling are among the approaches being explored.Source: §7.1.7 · p. 27
7.1.8 Scalability to Very Large Networks with Complex Constraints; 7.2.1 Developing Theoretical Frameworks for Fair IM; 7.2.3 Designing Highly Scalable Algorithms with Guarantees
- Fairness constraints can break submodularity and hinder scalability, motivating unified theoretical foundations and efficient approximation methods with guarantees for large networks.Proposed directions include axiomatic fairness frameworks, surrogate functions, reverse influence sampling, and unbiased estimators.Source: §7.1.8 · pp. 27–28, §7.2.1 · pp. 28–29, §7.2.3 · p. 29
7.1.9 Longitudinal Fairness and Long-Term Impacts; 7.2.2 Exploring Novel Fairness Notions suited for Network Dynamics
- Longitudinal fairness addresses cumulative effects of repeated interventions, while temporal, recurrent, and risk-aware notions target equity as networks and diffusion evolve.Repeatedly selecting easily reached influencers may concentrate social capital and leave other groups influence-deprived.Source: §7.1.9 · p. 28, §7.2.2 · p. 29
7.1.10 Bridging Algorithmic Fairness and Real-World Impact Assessment; 7.2.4 Improved Evaluation Methodologies and Benchmarks
- Real-world impact assessment asks whether algorithmic parity corresponds to social outcomes, and shared benchmarks are proposed to compare equity, efficiency, robustness, and context-specific performance.The survey identifies long-term behavioral change as a measurement challenge.Source: §7.1.10 · p. 28, §7.2.4 · pp. 29–30
7.2.5 Fairness in Complex Diffusion Models; 7.2.6 Interdisciplinary Research
- Future work extends fairness to complex, competitive, and topic-aware diffusion, while interdisciplinary research brings social-science insight and community feedback into model design.Hypergraphs or multilayer networks are proposed for complex interactions, and human-in-the-loop systems can incorporate qualitative feedback.Source: §7.2.5 · p. 30, §7.2.6 · p. 30
8 CONCLUDING REMARKS
The survey argues that influence maximization should balance spread with equitable impact and reports that DCDIM and CDIM best balance community coverage and disparity in its empirical analysis. It also identifies substantial gaps in scalability, contextual fairness, uncertainty, deployment, and interdisciplinary validation.
Concluding empirical findings
- DCDIM and CDIM emerge as the most effective algorithms for maximizing community coverage and minimizing disparity, while SSR-PEA, FIMMOGA, and PDTF fail to balance both.The survey reports that the latter approaches sacrifice influence spread to satisfy their specific fairness constraints.Source: §8 · pp. 30–31
Concluding remarks
- The survey emphasizes balancing influence performance with equitable impact, particularly where social, economic, or demographic disparities exist.Source: §8 · p. 30
- Open gaps include scalability, contextual fairness, uncertainty, real-world deployment, and interdisciplinary validation, alongside the need for robust, transparent, and interpretable models.Source: §8 · p. 31