Structure-Aware Cooperative Games

Updated 17 December 2025

Structure-Aware Cooperative Games are models that incorporate combinatorial and communication constraints into coalition formation and value allocation.
They extend classical TU games with adaptations like graph-based cores, dependency degrees, and operadic frameworks for structured interactions.
These frameworks enable advances in GNN explainability, resource allocation, and optimization, while introducing new computational challenges.

Structure-Aware Cooperative Games encompass a diverse and rapidly developing set of frameworks in which underlying combinatorial, communication, or dependency structures fundamentally shape coalition formation, value allocation, and solution concepts. Departing from the classical assumption of fully flexible coalition formation over 2^N, structure-aware models incorporate explicit constraints such as communication graphs, resource ownership partitions, physical coupling, or bounded dependencies, giving rise to new stability results, computational challenges, and interpretive tools across domains ranging from GNN explainability and combinatorial optimization to oligopoly and multi-objective resource allocation.

1. Formal Models and Structural Paradigms

Structure-aware cooperative games generalize the classical transferable-utility (TU) setting by endowing the coalition space and/or coalition values with additional combinatorial or algebraic structure:

Communication Graphs: Coalitions are feasible only if they induce connected subgraphs in a fixed undirected network ℰ over the player set N. This paradigm underpins the Hamiache–Navarro (HN) value, which explicitly propagates surplus only along edges, and graph-based cores where feasible sets are, e.g., the connected subgraphs of a tree or general graph (Zhang et al., 2022, Grabisch, 2013).
Dependency and Supermodular Degree: Interaction patterns are described by the dependency graph G_v, where edges denote dependency of one player's marginal contribution on another. The (supermodular) dependency degree parameterizes the tractability of solution concepts such as the Shapley value and the least-core, enabling FPT algorithms for games with bounded degree (Igarashi et al., 2017).
Partitioned Resource Structures: In Partitioned Combinatorial Optimization Games (PCOGs), each agent owns a subset of the ground set of a combinatorial problem (vertices/edges in a graph), and coalitional value is the optimum of the induced subproblem. This models agent-induced substructures and shifts complexity landscape compared to classical characteristic-function games (Chen et al., 25 Aug 2025).
Partition Function Games: In scenarios with externalities, values depend on the entire coalition structure or on embedded coalitions, which are pairs (S,π) consisting of a subset and a partition embedding. The lattice of embedded subsets brings a rich poset and Möbius inversion theory into solution concepts (Grabisch, 2010, Rossi, 2018).
Bayesian and Multi-Objective Structures: In cooperative Bayesian games, structural independence (agent/type) is represented by factor graphs, while in multi-objective resource allocation, structure-awareness is operationalized via Pareto-set inclusions between coalition structure refinements (Oliehoek et al., 2012, Pettersson et al., 27 Feb 2025).
Operadic and Algebraic Frameworks: The algebraic operad of cooperative games defines formal composition rules for aggregating subgames, encoding closure under structure-aware operations and compatibility of solution concepts (Mermoud et al., 19 Jun 2025).

2. Key Solution Concepts: Core and Value Extensions

Structure-awareness necessitates adapting classical solution concepts—core, Shapley value, marginal vectors, Weber set—to restricted coalition spaces or structured interactions.

Core on Feasible Set Systems: For games on a set system ℱ⊆2^N (e.g., all connected subgraphs, downsets in a poset), the core is defined by

$\Core(v,\mathcal F) = \{x \in \mathbb{R}^n : \sum_{i\in N} x_i = v(N),\ \sum_{i\in S} x_i \ge v(S)\ \forall S\in \mathcal F\},$

and its structure (boundedness, rays, faces) tightly depends on properties of ℱ such as union-closure, augmenting systems, and lattice regularity (Grabisch, 2013).

Graph-Restricted Values: The HN value $\phi_i(N,v,\mathcal{E})$ is the iterated limit value of singleton {i} in a surplus propagation process governed by the communication graph ℰ. Key properties include strict locality (zeroing out disconnected marginal contributions), edge-based iterative aggregation, and structural fidelity to message passing in GNNs (Zhang et al., 2022).
Möbius-Based and Chain Values: In poset/lattice-structured games, chain-based marginal vectors and Möbius inversion yield generalized Shapley values and Harasanyi dividends for settings with externalities or ordered feasibility (Grabisch, 2010, Rossi, 2018, Mermoud et al., 19 Jun 2025).
Overlap and Externality Cores: For overlapping coalition models (OCF-games), the c-core is defined over outcomes (coalition structures, resource splits, imputations) with generalized blocking constraints that incorporate resource divisibility and superadditive covers, with core-nonemptiness characterized by a suitable balancedness condition (Chalkiadakis et al., 2014).
Noncooperative Structure-Aware Equilibria: Structure-aware coalition formation via noncooperative games yields K-equilibria, where each player's strategy includes intended coalition participation and local actions, equilibrium involves a coalition-structure formation rule, and K parameterizes permitted simultaneous deviations (Levando, 2017).

3. Computational and Algorithmic Aspects

Structure-awareness both complicates and enables tractable computation depending on the interaction between game structure and solution concepts:

Tractability Via Parameter Boundedness: For games with bounded dependency or supermodular degree ( $d(v)$ , $p(v)$ ), FPT algorithms exist for computing Shapley/Banzhaf values and the least-core, exploiting the localization of significant marginal contributions (Igarashi et al., 2017).
PCOGs and Complexity Escalation: When agent resource-ownership partitions are large, core stability verification and existence problems exhibit Θ₂^P- or DP-completeness for graph optimization tasks such as minimum dominating set, compared to P-completeness for classical COGs. In contrast, for minimum spanning tree games, superadditivity yields a guaranteed nonempty core and polynomial algorithms (Chen et al., 25 Aug 2025).
Approximation Algorithms for Graph Games: WGGs (edge-weighted graph coalition games) exhibit NP-completeness for optimal coalition-structure, even on restrictive graph classes (planar, bounded degree); however, constant-factor approximation algorithms are enabled by algorithms that decompose positive-edge subgraphs into few forests or matchings (Bachrach et al., 2011).
Pareto Front Inclusion/Reduction in Multi-Objective PFGs: Exploiting the inclusion property of Pareto sets under coalition splitting, the computational cost of evaluating all coalition structures in multi-objective PFGs collapses to a single MOO plus post-processing via dominance checks—a highly scalable structure-aware method for games with large coalition spaces (Pettersson et al., 27 Feb 2025).
Factor Graphs in Bayesian Settings: For cooperative Bayesian games, factor-graph representations capturing agent and type independence underlie efficient solution methods (MAX-SUM message passing), breaking exponential dependence into polynomial complexity under bounded local scopes (Oliehoek et al., 2012).

4. Applications and Empirical Outcomes

Structure-aware cooperative game frameworks have yielded advances across several domains:

Model Explanation in GNNs: GStarX, a structure-aware explanation technique for GNNs, employs the HN value to score node importance respecting the intrinsic communication substructures. Empirically, GStarX achieves higher fidelity metrics and more intuitively structured explanations compared to Shapley-based baselines; e.g., recovering known motifs and semantic subgroups on molecular and sentiment datasets (Zhang et al., 2022).
Resource Allocation and Oligopoly: In cooperative oligopoly games with bounded rationality, structure-awareness is deployed via probabilistic beliefs over outsider coalition structures. This yields explicit coalition values as expectation over possible market partitions and strengthens core stability results as compared to pessimistic (γ-core) models (Lekeas et al., 2011, Lekeas et al., 2012).
Combinatorial Exchange and Scheduling: The PCOG framework models multi-organization kidney exchanges or scheduling pools, capturing how high-level resource partitioning among agents transforms both the analysis of core stability and the algorithmic complexity relative to combinatorial underpinnings (Chen et al., 25 Aug 2025).
Optimization in Subsurface Resources: Structure-aware MOO-PFGs provide a tractable framework for multi-agent collaboration in resources like water networks or CO₂ injection, addressing the intractability of brute-force coalition structure enumeration and enabling system-wide Pareto analysis (Pettersson et al., 27 Feb 2025).

5. Theoretical Insights and Generalizations

The study of structure-aware cooperative games has informed and unified various theoretical developments:

Operad Structure of Games: The algebra of all coalitional games is modeled as an operad, with composition and aggregation compatible with classical and structure-aware subspaces—additive, convex, balanced, totally monotone games—each forming suboperads. Solution concepts such as the core, Shapley value, and Banzhaf index behave functorially under operad composition (Mermoud et al., 19 Jun 2025).
Lattice and Poset Generalizations: The shift from Boolean lattices (2^N) to distributive, geometric, or embedded-coalition lattices expands both the meaning of convexity, balancedness, and the geometric structure of the core, providing tight characterizations of feasible allocations and marginal importance through Möbius calculus and chain decompositions (Grabisch, 2013, Grabisch, 2010, Rossi, 2018).
Externalities and Embedded Coalition Analysis: By describing solution concepts in the lattice of embedded coalitions, structure-aware models reveal the precise impact of externalities, establish equivalences and divergences between the classical Bondareva–Shapley balancedness and its partition-function analogues, and clarify which axioms and properties are preserved (Grabisch, 2010, Rossi, 2018).

6. Limitations, Open Problems, and Future Directions

Despite substantial progress, several challenges persist:

Scalability: Exact solution concepts (e.g., HN value, core enumeration) scale exponentially in the size of the underlying structure (e.g., node set, dependency degree), motivating sampling methods, FPT parameterizations, and approximation algorithms (Zhang et al., 2022, Igarashi et al., 2017).
Feature/Edge-Level Attribution: Most extant structure-aware value methods focus on node-level or coalition-level attributions. Extending these to edge-based or feature-attribution (e.g., the position value) remains an open trajectory, especially for complex models like transformers (Zhang et al., 2022).
Complexity under Generalized Structures: The impact of more intricate agent-resource partitions, high-degree dependency networks, or general submodular/convex set systems on tractability is incompletely classified and a focus of ongoing research (Chen et al., 25 Aug 2025, Bachrach et al., 2011, Igarashi et al., 2017).
Solution Concept Generalization: Broader deployment of operadic and lattice-theoretic tools for devising new solution concepts, refining fairness requirements, and executing efficient distributed computation in structure-aware environments is a contemporary research direction (Mermoud et al., 19 Jun 2025, Grabisch, 2010).
Empirical Validation and Domain Integration: Bridging theoretical advances to domain applications (multi-agent learning, market design, resource allocation) demands further development of empirically validated, domain-adaptable structure-aware solution methods.