Showing 38 papers for 2026-08-04
Federated Multimodal Graph Learning (FMGL) extends federated graph learning to multimodal attributed graphs, enabling collaborative optimization across decentralized MAGs without sharing raw data. However, naive FGL methods fail to navigate the multifaceted heterogeneity inherent in distributed MAGs, such as modality, topology, and data distribution, which hurts performance. This work proposes a principled approach to navigate these heterogeneities and improve FMGL effectiveness.
We introduce NLSD, a nonlinear Laplacian operator for signed and directed networks. It extends the signed Laplacian and directed Laplacian by computing node-specific potentials from node features and using a nonlinear potential discrepancy to drive diffusion. This operator improves learning on signed-directed graphs.
Compound flooding forecast in managed coastal systems is enhanced with multi-source dynamic graph learning that integrates hydrometeorological and operational signals from multiple monitoring stations. Single-site records fail to capture prolonged high-water plateaus, so distributed gage signals are fused to better represent high-water dynamics. The approach improves forecast accuracy and supports early warning.
LAB-Tab presents LLM-augmented Bayesian Network adaptation for few-shot tabular data generation. It addresses overfitting to tiny target statistics and the mismatch when reusing source-domain generators. The method blends target statistics with LLM-guided adaptation to produce plausible synthetic tabular data.
CoRe-GNN introduces multilevel message passing on coarsened graphs to reduce memory costs for large graphs. It analyzes limitations of coarsening methods—low-rank approximations may assign uniform representations, while Cluster-GCN restricts propagation intra-cluster—and offers a unified perspective on scalable propagation.
Weixin Pay-scale deep graph learning is applied to credit risk detection, tackling the scalability bottlenecks of processing heterogeneous, billions-scale user-risk graphs. The work demonstrates scalable GNN architectures that effectively mine risk signals for fraud detection in a massive digital payments ecosystem.
Benchmarking Sheaf Neural Networks for Inductive Tasks provides the first systematic evaluation of SNNs under inductive protocols. It compares three diffusion mechanisms and analyzes how edge-dependent restriction maps influence generalization to unseen graphs, guiding future design choices.
Pretraining on Call Graphs investigates how incorporating call-graph context improves binary function embeddings for analysis tasks. The study analyzes when call-graph-informed embeddings help code search, vulnerability detection, and malware classification, offering nuanced conclusions about contextual benefits.
Fairness in Augmented Graph Learning surveys challenges at the intersection of augmentation-based graph learning and fairness (termed FairGX). It develops a taxonomy of fairness concerns across federated learning, graph transformers, and graph condensation, and outlines potential mitigation strategies.
CountTRuCoLa introduces interpretable temporal KG forecasting with symbolic rules. It learns four simple rule types, including temporal rules with confidence that combines recency and frequency. Evaluated on nine datasets, it achieves competitive performance while keeping predictions directly traceable to the rules.
Rethinking Federated Graph Foundation Models proposes a graph-language alignment-based approach to federated graph foundation models. It argues that vector-quantized backbones cause irreversible knowledge loss and advocates semantic-structure alignment to improve generalization in federated settings.
GraphER proposes an efficient graph-based enrichment and reranking method for retrieval-augmented generation. It builds a graph over evidence items to enrich and rerank retrieved content, reducing the need for iterative agentic retrieval and lowering computational overhead.
K-STEMIT introduces a knowledge-informed spatio-temporal multi-branch GNN for subsurface stratigraphy thickness estimation from radar data. It leverages domain knowledge to mitigate noise and artifacts in radargrams and combines spatial, temporal, and knowledge-driven cues for improved thickness estimation.
Dimensionality Reduction Meets Network Science examines UMAP's internal kNN graph, showing that the graph preserves the data manifold before projection. It demonstrates using standard graph algorithms on this graph (e.g., PageRank) to enhance sensemaking, such as identifying representative points and structure.
Learned, Relied Upon, or Necessary? Separating Checkpoint Dependence from Task-Level Value in Sheaf GNNs disentangles whether learned restriction maps truly help or merely reflect a checkpoint's organization. The paper introduces estimands to separate checkpoint reliance from genuine map utility using retraining and controlled ablations.
Improved convergence rate of kNN graph Laplacians introduces a differentiable self-tuned affinity where the graph weights depend on local density estimates. This adaptive bandwidth improves convergence properties of the Laplacian and enhances manifold learning on data with variable density.
The Graph Language introduces GRALAN, a framework that lets knowledge graphs speak to large language models via relational tokens. A trainable language mediator converts KG structure into tokens compatible with frozen LLMs, enabling grounding for knowledge-intensive tasks. Demonstrated on QA tasks framed as entity classification.
G-ReAct proposes graph-guided deep search with structure-state co-evolution for open-domain tasks. It preserves intermediate states and constraints during long-horizon multi-hop search, reducing context forgetting and search drift, and enables more efficient exploration by coordinating graph structure with reasoning state.
GABench is a comprehensive benchmark for evaluating LLM agents on graph analysis tasks, covering diverse graph types and tasks to systematically assess agent capabilities. It aims to standardize evaluation and drive progress in graph-enabled LLM agents.
CockpitHAT introduces a dependency-graph-driven hierarchical attribution framework for embodied multi-agent cockpits. It addresses Correctness Collapse by leveraging dependency structure and multi-channel evidence for safety-aware attribution, replacing simple positional windows with dependency-based analysis.
CoEvoKG introduces a co-evolution framework where a knowledge graph powers both task generation and a persistent evidence memory for agents. The task generator creates multi-hop questions from entity chains in the KG, while the search agent learns from solving those tasks and the memory preserves successful search traces to guide future learning.
HPFA proposes a hypergraph-based paired failure attribution framework to attribute LLM reasoning failures to specific steps, addressing non-linear dependencies and avoiding expensive counterfactual testing. By organizing failures on a hypergraph in paired form, it aims to identify root causes and guide targeted improvements.
SkillTrace builds a three-level query-skill graph to enable composable LLM agents. It models relations among skill queries, similarity between queries and library skills, and dependencies among selected skills to discover complete, executable skill compositions.
KING proposes Embodiment-Aware Kinematic Graph Neural Network to learn unified motion representations across legged and wheeled robots, improving odometry in feature-poor environments by capturing nonlinear dynamics. The model leverages a kinematic graph approach to align proprioceptive signals across embodiments.
SG-Layout introduces a graph-guided layout generation framework that injects structured spatial knowledge into LLMs. It uses a two-stage training: first a graph-language feature alignment between a relational scene graph and language, then leveraging the aligned features for LLM-based layout generation that respects explicit geometric relations.
X-KGRank presents a knowledge graph retrieval-augmented framework for explainable recommendations by combining pattern mining and LLM re-ranking. It unifies structural collaborative filtering with LLM-based explanations, grounding recommendations in the KG and using mined patterns to re-rank candidates.
TBSG-Net introduces Temporal Bipartite Scene Graph Network for fine-grained video moment retrieval. By modeling objects and relations as temporally evolving scene graphs, it captures temporal dynamics and defines explicit temporal spans to improve localization.
Knowledge Graph Augmented LLMs for Disease Prediction proposes a KG-guided chain-of-thought framework for visit-level prediction on MIMIC-III. It maps ICD-9 codes to PrimeKG, mines disease-relevant nodes and paths to scaffold temporally consistent CoT rationales, and fine-tunes lightweight LLMs.
Panning for Gold: Expanding Domain-Specific Knowledge Graphs with General Knowledge explores how general knowledge graphs can enrich domain-specific KGs by systematically leveraging high-quality GKGs to broaden coverage, improve reasoning, and integrate external knowledge.
OntoTKGE proposes ontology-enhanced temporal knowledge graph extrapolation, using ontological knowledge to alleviate sparse histories by enabling entities to inherit behavioral patterns from same-concept entities. An encoder-decoder framework models temporal evolution with ontology-informed priors.
Unifying biomedical knowledge in a modern multimodal graph introduces OptimusKG, a multimodal biomedical labeled property graph built from structured and semi-structured resources to preserve factual, type-specific metadata across molecular, anatomical, clinical, and environmental domains.
The Violation Situation Pattern proposes a reusable ontology pattern to persistently represent compliance violations as graph entities with a rule identifier, temporal validity, lifecycle state, and evidence links, enabling audit and traceability.
Representing Higher-Order Networks: A Survey of Graph-Based Frameworks surveys higher-order graph formalisms that extend beyond pairwise edges to multiway, hierarchical, temporal, multilayer, recursive, and tensor-based interactions, with discussion of models and applications.
LongCrafter proposes a structured synthesis framework that combines a hierarchical task taxonomy with an evidence-grounded generation pipeline to create diverse, difficult, and faithful long-context SFT data for LLMs.
GARDRec introduces decision-level graph grounding to constrain LLM-based recommendations with structured user-item relations, improving ranking fidelity under same-context comparisons.
HyperAgent4POI uses Dynamic Semantic Message Passing on multi-agent hypergraphs to perform modality completion and refine incidences, enabling better POI recommendations when textual/visual content is missing.
Multimodal Embeddings for 3D Similarity Search in Semantic Web-of-Things Digital-Twin Platforms develops multimodal embeddings to support 3D similarity search, enabling queries that combine ontological constraints with cross-modal similarity over temporally evolving scene data.
MERIT presents efficient in-place deletion techniques for dynamic graph-based ANNS indexes, enabling removal of obsolete vectors without interrupting online queries or requiring expensive graph-wide maintenance.