Global Preference Graph Overview

• Global preference graphs are combinatorial structures that encode distributed preference information across agents using graph connectivity.
• They enable algorithmic analysis and optimization in voting, recommendation systems, and game-theoretic dynamics with clear complexity insights.
• Applications include detecting consensus in elections, enhancing recommendation accuracy, and analyzing strategic interactions in finite games.

A global preference graph is a combinatorial or algebraic structure that captures collective or distributed preference information across populations, sessions, agents, or players over a common set of alternatives. In its most abstract form, it encodes how preferences, rankings, or choices interact or agree within a set of agents, sessions, or game profiles by exploiting an underlying graph structure. This framework supports algorithmic analysis of voting, recommendation, distributed agreement, sequential choice, and game-theoretic dynamics. This entry reviews key definitions, optimization principles, structural theorems, complexity findings, recognition algorithms, and applications, as well as comparing leading instantiations and interpretations.

1. Formal Definitions and Model Classes

The global preference graph is instantiated differently across literature depending on the application, but always encodes a structured aggregation of preference data:

• Single-Peakedness and Traversal (Voting/Preference Aggregation): Let $C = \{1,\dots,m\}$ be candidates and $P = \{R_1,\dots,R_n\}$ orderings. A connected undirected graph $G = (C, E)$ is a global preference graph if every ranking in $P$ "traverses" $G$: for each $i$, every $k$, the subgraph induced by the top-$k$ in $R_i$ is connected. The sparsest such $G$ reflects the shared backbone structure of the electorate (Escoffier et al., 2020).
• Session-based and Recommender Systems: Global graphs typically combine multiple layers or relations, e.g., transitions across all user sessions, co-occurrence relations, and user-item interactions. In heterogeneous variants, the global preference graph is a typed, often directed graph $(V, E, R)$ with items, users, sessions, and multiple edge types, encoding item–item transitions, global co-occurrences, and user–item history (Pang et al., 2021, Wang et al., 2021, Ding et al., 2021, Lim et al., 2020).
• Networked Preference Dynamics: Each agent $i$ updates a preference relation $\pi_i$ over alternatives $A$, communicating via a network $G = (N,E)$. The system's global preference profile evolves under distributed message passing and lattice-theoretic aggregation, forming a global preference graph over $\operatorname{Pre}(A)^N$ (Riess et al., 2023).
• Game Theory: In finite games, the preference graph $G = (V, E)$ has vertices as strategy profiles; an edge $(p, q)$ is drawn if $q$ differs from $p$ by one player's strategy and yields higher utility for that player. The entire graph captures the global structure of incentives and dynamics (Biggar et al., 5 Feb 2025).

The general schema is thus: nodes represent alternatives, agents, items, or profiles; edges capture pairwise or higher-order relations derived from preferences or behavior aggregated at the collective or network level.

2. Optimization and Complexity: Constructing Global Preference Graphs

In the single-peakedness model, the central problem is to construct a global preference graph $G$ of minimal complexity compatible with a given profile $P$:

• ILP Formulations: Variables $x_{kl} \in \{0,1\}$ indicate edges. Two main objectives have been studied (Escoffier et al., 2020):
  • Minimize $|E|$: $\min \sum_{k<l} x_{kl}$ subject to traversal constraints.
  • Minimize maximum vertex degree $\Delta(G)$ via auxiliary variables and degree constraints.
• NP-Hardness: Both minimal edge and minimal degree global preference graph construction are NP-hard via reduction from Set-Cover (Escoffier et al., 2020).
• Polynomial Algorithms in Special Cases: If the optimal number of edges is $m-1$, corresponding to tree solutions, the ILP relaxes to an LP whose solution is always integral, enabling polynomial-time recognition for trees (and, with additional constraints, paths and pseudotrees) (Escoffier et al., 2020).

In recommender systems, global preference graphs are typically not optimized for sparsity but designed to maximize information propagation and context aggregation, often via k-hop message passing on large but moderately sparse graphs (Wang et al., 2021, Pang et al., 2021, Ding et al., 2021).

3. Recognition, Learning, and Message Passing

Recognition and learning in global preference graphs divides into algorithmic recognition (compatibility detection) and neural or lattice-based message passing.

• Recognition Algorithms (Voting): For paths (classical axes), there exist $O(nm)$ algorithms incrementally building an ordering consistent with interval-top sets. For trees and pseudotrees, LP-based and combinatorial algorithms are provided, leveraging tight traversal constraints and flow arguments (Escoffier et al., 2020).
• Message Passing (Networked Agents): Agents update preferences $\pi_i$ via lattice-based aggregation, e.g., median or join of neighbors’ preferences. The system forms a global dynamical system $P(t+1) = F(P(t))$, with monotonicity and lattice structure guaranteeing fixed points and, for inflationary maps, convergence in finite time (Riess et al., 2023).
• Neural Propagation (Recommendations): Heterogeneous GNNs pass messages over typed global graphs using relation-specific layers. Session-aware attention and global-to-local fusion yield item and user embeddings incorporating both immediate behavior and long-range context (Wang et al., 2021, Pang et al., 2021, Ding et al., 2021, Lim et al., 2020).

Application Domain	Recognition/Training Principle	Graph Structure
Voting (single-peaked)	ILP/LP; combinatorics	Minimal connected graph
Agent Networks	Lattice-theoretic aggregation	Network of agents
Recommender systems	GNN; attention; LightGCN	Multi-relational, global
Game theory	Graph-theoretic analysis	Strategy profile graph

4. Structural Theorems and Interpretations

Structural results for global preference graphs connect them to fundamental properties of underlying systems.

• Game Theory: The preference graph characterizes dominated strategies (out-degree in all opponent contexts), pure Nash equilibria (sinks of $G$), potential games (acyclic $G$), supermodularity (row/column monotonicity in edge orientation), and weak acyclicity (sink-components in best-response subgraph are singletons) (Biggar et al., 5 Feb 2025).
• Consensus and Branching: In voting/global ranking, graph sparsity ($|E|$ minimal) encodes strong consensus, while high maximum degree $\Delta$ signals the presence and extent of branching within the underlying consensus axis (Escoffier et al., 2020).
• Recommendation Filtering: In session-based systems, edges in the global graph represent cross-session pairwise transitions; attention mechanisms can dynamically filter global context to favor items most relevant to a current session (Wang et al., 2021).
• Lattice Dynamics: Monotonicity and completeness of the global preference profile's lattice structure guarantee the existence and convergence of equilibrium profiles, with extremal equilibria corresponding to maximal/minimal consensus (Riess et al., 2023).

5. Applications and Empirical Findings

Global preference graphs underpin a variety of applications in computational social choice, recommendation, distributed consensus, and game dynamics.

• Recommendation Systems: Joint modeling of user-item and item-item global graphs using GNN enhances sequential or session-based recommendation accuracy by propagating high-order context signals and alleviating data sparsity (Ding et al., 2021, Wang et al., 2021, Pang et al., 2021, Lim et al., 2020). Empirically, inclusion of global graphs consistently improves hit-rate and mean reciprocal rank on real datasets.
• Voting and Societal Preferences: The discovery (or impossibility) of a sparse global graph structure reveals the extent of shared preference structure, guides visualization, and enables efficient recognition of single-peakedness in large electorates (Escoffier et al., 2020).
• Network Preference Dynamics: The global preference lattice system enables algorithmic analysis of distributed preference evolution, with provable convergence and tunable dissensus or consensus via choice of local aggregation rules (Riess et al., 2023).
• Game Theory and Dynamics: The structure of the preference graph determines dynamic properties (e.g., fictitious play, replicator convergence) as well as equilibrium structure (e.g., uniqueness, location, and basins of attraction for Nash equilibria) (Biggar et al., 5 Feb 2025).
• Empirical Complexity: Solving the global min-edge ILP on real voting datasets with up to $m = 30$ and $n = 5000$ is practical (sub-second per instance); real-world domains manifest a diversity of global preference-graph topologies, from trees to complete graphs depending on the preference profile (Escoffier et al., 2020).

6. Synthesis and Theoretical Significance

The global preference graph formalism unifies disparate approaches to collective preference modeling by focusing on the minimal or otherwise expressive structure supporting observed rankings, behavioral traces, or strategic incentives. Its versatility underlies key complexity results (NP-hardness outside narrow graph classes), rigorous connections with core game-theoretic notions (dominance, equilibrium, potential), and recent advances in neural and distributed learning over large-scale, heterogeneous, graph-structured preference data. The global preference graph thereby offers both a stringent test of preference structure (how much consensus or diversity truly exists) and a flexible substrate for downstream inference, recommendation, distributed learning, and strategic analysis (Escoffier et al., 2020, Wang et al., 2021, Riess et al., 2023, Biggar et al., 5 Feb 2025).