← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 26 papers for 2026-05-14

Towards Robust Federated Multimodal Graph Learning under Modality Heterogeneity
GNN Graph Learning

This work addresses robust federated multimodal graph learning under modality heterogeneity. Real-world graphs are often siloed and modalities incomplete, which challenges centralized MGL methods. The authors propose a federated framework to integrate heterogeneous modalities without sharing raw data and strategies to handle missing modalities while preserving performance.

Graph-Based Financial Fraud Detection with Calibrated Risk Scoring and Structural Regularization
GNN Graph Learning

Financial transaction fraud is difficult due to complex relationship structures and evolving data distributions. The paper proposes a graph neural network representation learning and risk discrimination framework for fraud prevention that fuses transaction records and identity information to model group collaborations and chain transfers. The framework includes calibrated risk scoring and structural regularization to capture network-level patterns.

\emph{DRIFT}: A Benchmark for Task-Free Continual Graph Learning with Continuous Distribution Shifts
GNN Graph Learning

DRIFT introduces a benchmark for task-free continual graph learning under continuous distribution shifts. It relaxes the assumption of predefined task boundaries to reflect realistic non-stationary environments. The benchmark provides evaluation protocols and baselines to assess performance under gradual shifts.

Rethinking Efficient Graph Coarsening via a Non-Selfishness Principle
Graph Theory Graph Learning

This paper rethinks graph coarsening with a non-selfishness principle. Moving away from pairwise selfish matching, the method promotes collective information gathering to reduce computation and memory overhead. The approach better preserves essential graph structure while scaling to large graphs.

What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition
GNN Graph Learning

This work introduces Tri-Component Information Decomposition (TIDE) for out-of-distribution detection on graphs. It explicitly decomposes information into feature, structural, and interaction components to identify what drives predictions under shifts. The framework improves robustness to shifts in node features and graph structure.

MLGIB: Multi-Label Graph Information Bottleneck for Expressive and Robust Message Passing
GNN Graph Learning

MLGIB tackles over-squashing in deep multi-label GNNs. In multi-label graphs, neighbor labels can be sparse and noisy, diluting useful signals. Multi-Label Graph Information Bottleneck filters irrelevant label information during message passing to improve expressivity and robustness.

Beyond Oversquashing: Understanding Signal Propagation in GNNs Via Observables
GNN Graph Learning

Beyond Oversquashing paper proposes an observables-based perspective on signal propagation in GNNs inspired by quantum mechanics. It defines signal-observables to track routing and diffusion of information across the graph. The work provides insights to mitigate information loss and guide architecture design.

Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models
Graph Learning

This work introduces decoupled and divergence-conditioned prompting for multi-domain dynamic graph foundation models. The prompts separate domain-specific knowledge from shared knowledge and condition prompting on divergence to reduce negative transfer. The approach improves generalization across domains in dynamic graphs.

Rethinking Generalization in Graph Neural Networks: A Structural Complexity Perspective
GNN Graph Learning

The paper studies generalization of GNNs from a structural complexity perspective. It argues that graph structure, not only model capacity, governs generalization due to topological dependencies. The authors develop a structural complexity framework with theoretical and empirical insights.

Multimodal Graph-based Classification of Esophageal Motility Disorders
GNN Graph Learning

The work explores multimodal graph-based classification of esophageal motility disorders. It combines HRIM recordings with patient information in a graph-based model to classify disorders. The study analyzes data from 104 patients to demonstrate feasibility.

Graph Neural Networks with Triangle-Based Messages for the Multicut Problem
GNN Graph Learning

The paper introduces triangle-based messages for GNNs in the multicut problem. It uses edge-centric features and triangle-based message passing to encode local cycle constraints. Experiments on synthetic and real-world data show effectiveness.

The WidthWall: A Strict Expressivity Hierarchy for Hypergraph Neural Networks
GNN Graph Learning

The WidthWall establishes a strict expressivity hierarchy for hypergraph neural networks. Expressivity is characterized by the small patterns architectures can detect and count via homomorphism densities. The framework integrates classical and hypergraph-centric notions to map architecture capabilities.

GraphIP-Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It?
GNN Graph Learning

GraphIP-Bench provides a unified benchmark for GNN model extraction attacks and defenses. It standardizes threat models, datasets, and metrics to compare how hard it is to steal a GNN and how defenses fare. The benchmark enables fair cross-study evaluation of ownership defenses.

Gradual Domain Adaptation for Graph Learning
Graph Learning

Gradual Domain Adaptation for Graph Learning (GGDA) constructs a sequence of intermediate graphs bridging source and target domains. It uses Fused Gromov-Wasserstein to generate knowledge-preserving intermediates that minimize information loss, enabling stable adaptation under large distribution shifts.

Exact Verification of Graph Neural Networks with Incremental Constraint Solving
GNN Graph Learning

Exact Verification of GNNs with Incremental Constraint Solving presents an exact robustness verifier. It certifies against attribute and structural perturbations including edge addition/deletion for common aggregation functions. Incremental constraint solving makes the approach scalable to practical networks.

Learning to Approximate Uniform Facility Location via Graph Neural Networks
GNN Graph Learning

Learning to Approximate Uniform Facility Location via Graph Neural Networks studies GNN-based heuristics for UniFL. It tackles the trade-off between differentiable learning methods and classical worst-case guarantees. The paper demonstrates performance on UniFL benchmarks and discusses guarantees.

RAG-GNN: Integrating Retrieved Knowledge with Graph Neural Networks for Precision Medicine
GNN Graph Learning

RAG-GNN integrates retrieved knowledge with GNNs for precision medicine. It is an end-to-end retriever-augmented framework with retrieval projection, gated fusion, and contrastive alignment. In a cancer signaling case study it improves functional clustering and modeling.

IdeaForge: A Knowledge Graph-Grounded Multi-Agent Framework for Cross-Methodology Innovation Analysis and Patent Claim Generation
Knowledge Graph

IdeaForge presents a knowledge graph-grounded multi-agent framework for cross-methodology innovation analysis. It combines multiple methodologies (e.g., TRIZ, Design Thinking) and uses a KG to coordinate reasoning. It supports patent claim generation with traceable, cross-method reasoning.

Representing Higher-Order Networks: A Survey of Graph-Based Frameworks
Graph Theory

A comprehensive survey of higher-order network representations beyond pairwise graphs. It covers multiway, hierarchical, temporal, multilayer, recursive, and tensor-based models. It discusses mathematical foundations and practical implications for representing complex systems.

Beyond Individual Mimicry: Constructing Human-Like Social network with Graph-Augmented LLM Agents
Graph Learning

GraphMind proposes graph-augmented LLM agents to form human-like social networks. It equips LLM-driven bots with graph-awareness to coordinate over global network structure. The framework aims to enhance realism and resilience against graph-based bot-detection.

GAAMA: Graph Augmented Associative Memory for Agents
Graph Learning

GAAMA introduces a graph-augmented associative memory for agents, constructing a concept-mediated knowledge graph to preserve the structural relationships among memories across sessions, addressing issues in flat retrieval-augmented generation and mega-hub effects in entity-centric graphs. It employs a three-stage process to integrate memory into agent behavior for persistent, personalized interactions.

T-TExTS (Teaching Text Expansion for Teacher Scaffolding): Enhancing Text Selection in High School Literature through Knowledge Graph-Based Recommendation
Knowledge Graph

T-TExTS is a knowledge-graph–based recommendation system that selects high school literature texts based on pedagogical merit rather than surface metadata. It builds a domain-specific ontology using KNARM and instantiates it to support text expansion that scaffolds teaching.

Linking Extreme Discourse to Structural Polarization in Signed Interaction Networks
Graph Theory

This work presents a language-grounded signed-network pipeline that derives continuous signed edge weights from LLM stance scores to link discourse with interaction structure, enabling a unified measurement of polarization. It quantifies how extreme discourse relates to structural polarization in signed networks.

When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method
LLM × Graph

The paper investigates how prompt design, cultural framing, language, and model scale affect LLM-generated realistic social networks. It formalizes four tie-formation mechanisms—sequential, global, local, and iterative—as distinct conditional distributions over edges, evaluated using a fixed roster of 50 demographic groups.

Graph-Grounded Optimization: Rao-Family Metaheuristics, Classical OR, and SLM-Driven Formulation over Knowledge Graphs
Knowledge Graph

Graph-grounded optimization introduces a paradigm where decision variables, constraints, and objective coefficients are sourced from a knowledge graph via Cypher queries. It situates this approach among recent LLM/SLM-driven optimization systems and demonstrates instantiations using Rao-Family metaheuristics and SLM-driven formulations for real-world problems.

SemRepo: A Knowledge Graph for Research Software and Its Scholarly Ecosystem
Knowledge Graph

SemRepo is an RDF knowledge graph with over 81 million triples describing nearly 200,000 GitHub repositories tied to scientific research. It captures repository metadata such as contributors, issues, and languages, and interlinks authors with SemOpenAlex, repositories with publications in LPWC, and research artifacts like datasets and experiments to form a connected scholarly software ecosystem.