SWAN: Social Welfare & Network Effects

• SWAN is a framework that maximizes aggregate social welfare by incorporating individual-specific preferences and network-induced externalities.
• It employs mathematical, spectral, and greedy algorithmic approaches to efficiently optimize interventions under budget and fairness constraints.
• The framework has practical applications in public policy, federated learning, and viral marketing, demonstrating significant improvements in welfare and cost efficiency.

Social Welfare Maximization with Application-Aware and Network Effects (SWAN) encompasses a set of mathematical, algorithmic, and mechanism-design frameworks for maximizing aggregate utility (social welfare) in networks, explicitly accounting for both application- or agent-specific intrinsic preferences (“application awareness”) and the interdependencies introduced by network (externality) effects. SWAN-formulated problems arise in domains including public policy, networked market design, federated learning, information diffusion, and targeted interventions, where the planner must optimally allocate incentives, treatments, or resources while respecting the interplay between intrinsic agent characteristics and their position or interactions within a network. This entry surveys foundational models, algorithmic paradigms, solution approaches for both continuous and discrete SWAN formulations, mechanism design aspects, and empirical findings.

1. Problem Formulations and Core Principles

At its core, the SWAN paradigm specifies a social welfare objective over a networked population, where each agent’s payoff is a function of (i) individual-specific features or preferences relevant to the application (“application awareness”), and (ii) network effects, typically modeled via adjacency matrices or propagation probabilities capturing how an agent's utility is influenced by the actions or states of their neighbors.

Several canonical mathematical instantiations exist:

• Linear–Quadratic Network Games: Agents $i\in\mathcal{N}$ choose actions $x_i$; their utilities depend on a baseline vector $b$ (application features) and a symmetric weighted adjacency matrix $A$ (network structure), with spillover parameter $\phi$ capturing the direction and strength of network effects. The planner may intervene jointly in $b$ and $A$ (the “joint-intervention” model) under a quadratic budget constraint, maximizing $W(b, A) = b^\top (I - \phi A)^{-2} b$ after intervention costs (Kor et al., 2022).
• Allocation with Externalities: Agents and a set of items/applications, with personalized utility functions $h_j(\phi_i)$ depending on item and agent features, plus a network-externality term $a^i_{jk}$ for pairwise effects, yield an aggregate welfare objective incorporating both intrinsic and inter-agent terms. The problem is to allocate items to agents to maximize this sum under assignment constraints (Etesami, 2021).
• Utility-driven Influence/Adoption Models: In influence maximization and competitive propagation, agents may adopt bundles of items based on utility functions $U(A)$ that combine valuation, price, and random taste shocks, propagated via stochastic diffusion models (e.g., UIC: Utility-Driven Independent Cascade) (Banerjee et al., 2018, Banerjee et al., 2020).
• Mechanism Design for Federated Learning: Here, social welfare $W$ depends on aggregate model utility $U(\epsilon)$ (linked to generalization error and hence the set of participating clients), participation costs, and application performance constraints, underpinned by non-monotonic network effects on model quality and incentives (Li et al., 8 Jan 2026).

The SWAN class thus generalizes classical influence maximization, welfare optimization, and targeted intervention frameworks to settings where both agent-specific attributes and network-mediated externalities crucially affect optimal policy.

2. Analytical and Algorithmic Frameworks

The solution paradigms for SWAN problems depend heavily on the analytic structure of the welfare function, the type of externalities, and the nature of agent-level heterogeneity.

Spectral and Continuous Optimization Approaches

When the objective admits a continuous relaxation (e.g., joint intervention in linear–quadratic network games), spectral decomposition yields closed-form first-order conditions for optimal budget allocation:

• Spectral Characterization: In the moderate-budget regime, optimal link modifications $\Delta A_{ij}$ are proportional to products of the principal eigenvector's components, i.e., centralities $u^1_i u^1_j$, where $u^1$ is the leading eigenvector of the (post-intervention) network. For large budgets, the optimal intervention transitions to the extreme network structures: complete graphs (for complements) or complete balanced bipartite graphs (for substitutes), maximizing or minimizing the principal eigenvalue respectively (Kor et al., 2022).
• Lovász and Multilinear Extensions: For discrete allocation settings with (sub-/super-)modular structure induced by externality curvature, the Lovász extension (convex relaxation) and multilinear extension (for maximization) enable the application of continuous (fractional) relaxation and rounding methods, yielding constant-factor approximation guarantees. The continuous-greedy algorithm is particularly effective when the welfare is monotone submodular (Etesami, 2021).

Greedy and Sample-based Algorithms

For influence maximization and utility-driven diffusion models where submodularity is lost (due to complementarity or competition):

• bundleGRD and PRIMA: Prefix-preserving greedy routines are deployed for seed selection in large networks, with theoretical $(1-1/e-\epsilon)$-approximation guarantees in specific regimes. RR-set sampling underpins scalability for millions of nodes. When objective functions are not monotone or submodular, sequential or utility-guided greedy algorithms (SeqGRD) and block-wise accounting techniques are used to bound welfare loss (Banerjee et al., 2018, Banerjee et al., 2020).
• Welfare-based Fairness Optimization: Aggregating community-level utilities through isoelastic social welfare functions parametrized by an inequality-aversion index ($\alpha$) interpolates between efficiency and equity objectives and is amenable to monotone submodular maximization via greedy algorithms (Rahmattalabi et al., 2020).

Policy Targeting with Spillovers

In experimental or quasi-experimental settings with unobserved or partially observed networks, efficient policy targeting is achieved via:

• Semiparametric Welfare Estimation: Augmented inverse propensity score methods and network cross-fitting handle interference and overlapping neighborhoods, providing unbiased estimation of sample-analog welfare under interference (Viviano, 2019).
• MILP Formulations: Exact maximization of the estimated welfare objective (incorporating observed application constraints) is posed as a mixed-integer linear program, tractable for realistic instances and extendable to incorporate fairness, budget, or grouping constraints.

3. Mechanism Design and Incentive Compatibility

A critical aspect of SWAN in engineered networks (e.g., FL, crowdsourced data collection) involves designing mechanisms that both maximize welfare and ensure incentive-compatibility, budget-balance, and individual rationality under application performance constraints.

• MoTS and SWAN Mechanisms for Federated Learning: Clients choose among join (participate in training), buy (obtain model via purchase), or abstain. The mechanism sets participation rewards $r_i$ and model price $p$, funded by buyer payments, balancing participation incentives through a surplus-redistribution parameter $\tau$. The SWAN mechanism guarantees that the welfare-maximizing equilibrium is implemented as a unique Nash equilibrium, with proven gains in welfare and dramatic reductions in extra incentive costs compared to standard mechanisms (Li et al., 8 Jan 2026).
• Allocation, Budget-Balance, and Truthfulness: SWAN implements truthful reporting and budget balance as its payment rules depend only on induced social states from the clients' true types and satisfy stringent application error constraints. The solution adjusts to network-induced non-monotonicities in marginal benefits, temporarily over-incentivizing some classes (e.g., low-data clients) when needed to escape negative externality regions.

4. Welfare–Equity Trade-offs and Distributional Effects

SWAN formulations clarify the tension between maximizing total utility and ensuring equitable outcomes:

• Spectral Inequality Analysis: As budget increases in joint-intervention models, the dispersion of agent payoffs (measured by the Theil T-index) converges to minimal levels (even zero) in networks displaying perfect symmetry (i.e., complete graphs under complements). However, at moderate budgets (with nonuniform centralities), interventions may transiently exacerbate inequality before the optimal structure is achieved (Kor et al., 2022).
• Fairness Tuning via Social Welfare Functions: Isoelastic formulations allow explicit control over the welfare–equity trade-off by modulating the inequality aversion parameter. Empirical studies (e.g., on disaster preparedness in partitioned communities) demonstrate that moderate values of $\alpha$ reduce inter-community utility gaps with only minor efficiency loss, offering practical levers for equity-aware policy design (Rahmattalabi et al., 2020).

5. Empirical Findings and Practical Implementations

Key empirical evidence substantiates the robustness and scalability of SWAN methodologies across diverse domains:

• Federated Learning Prototypes: Hardware deployments (e.g., 20 Raspberry Pis on three canonical datasets) demonstrate that SWAN mechanisms can yield up to 352% improvements in welfare and 93% reductions in incentive cost relative to classical FL paradigms, with robust handling of heterogeneous client populations and arbitrarily strict application accuracy guarantees (Li et al., 8 Jan 2026).
• Large-Scale Network Diffusion: On real social networks with millions of nodes, welfare-maximizing seed allocation algorithms scale efficiently and consistently outperform adoption-maximizing or naive greedy baselines, with welfare gains persisting in both competitive and complementary item regimes (Banerjee et al., 2018, Banerjee et al., 2020).
• Experimental Policy Targeting: In policy targeting under network interference (e.g., insurance uptake in agrarian networks), SWAN-optimal individualized treatment rules leveraging spillover structure attain up to 30 percentage-point improvements in policy-relevant take-up rates compared to spillover-oblivious benchmarks (Viviano, 2019).

6. Unified Frameworks and Theoretical Guarantees

Recent advances provide unifying frameworks subsuming classical models from influence maximization, resource allocation under externalities, and utility-driven agent behavior:

• Unified Submodular Optimization: By explicitly separating application-aware linear terms and modeling network effects via parametric externality functions (linear, concave, convex), a single analytic scheme captures all tractable regimes of the SWAN problem. Matching relaxations (Lovász, multilinear) and tailored rounding algorithms yield the best-known (or improved) approximation ratios and allow seamless integration of application-level complexities and hard constraints (Etesami, 2021).
• Optimality and Regret Bounds: SWAN mechanisms and estimators admit rigorous performance guarantees: in continuous allocation, optimality is characterized spectrally; in semiparametric estimation, minimax regret bounds hold; in algorithmic allocation, constant-factor approximation is achieved whenever sub-(super-)modularity or fractional relaxations are available (Kor et al., 2022, Etesami, 2021, Viviano, 2019).
• Extensibility to Constraints: Both theoretical and empirical works confirm that SWAN formulations support the inclusion of application-specific constraints such as fairness (group, max-min), capacity limits, knapsack/matroid constraints, and latency/resource restrictions, with minimal impact on the underlying tractability or approximation guarantees.

7. Applications and Future Directions

SWAN approaches are actively deployed and researched in diverse application areas, including:

• Public Health and Information Campaigns: Fair and efficient seeding strategies maximize risk-weighted benefit across structure-aware communities (Rahmattalabi et al., 2020).
• Viral Marketing and Platform Design: Maximizing user engagement and aggregate platform utility via welfare-centric, network-aware marketing, and recommendation policies (Banerjee et al., 2018, Banerjee et al., 2020).
• Federated Learning and Decentralized Training: Mechanism design for distributed learning under network externalities, tailored to heterogeneous stakeholders and strict application-level performance (Li et al., 8 Jan 2026).
• Targeted Interventions in Policy Networks: Allocation under network interference with data-driven, robust welfare estimation and policy targeting (Viviano, 2019).

Future research directions include the systematic study of dynamic and time-evolving SWAN settings, learning-based approaches for estimating externalities and agent utility models in nonstationary networks, and advanced mechanism design for multi-objective, budgeted interventions under imperfect or adversarial information.