Showing 34 papers for 2026-05-21
In the transductive setting where the full graph is observed but node labels are partially available, this work notes that predictions for unlabeled nodes can carry useful training signal that standard supervised objectives ignore. It proposes a training objective that leverages these unlabeled-node predictions to sharpen learning and improve semi-supervised node classification.
WaveGraphNet introduces a coupled inverse–forward graph learning framework with physics constraints to localize damage from guided-wave measurements in composite plates. By enforcing physics-consistency and jointly learning inverse mappings and forward simulations, it improves localization accuracy and generalization to unseen damage regions with sparse sensor networks.
TriForces presents a model-agnostic three-stream framework that separates composition and structure information to create more transferable representations for interatomic potentials. By preserving both compositional and structural context, TriForces enhances cross-domain transfer when adapting MLIPs to new chemistries with limited data.
Graph Navier Stokes Networks (GNSN) extend graph neural networks by incorporating convection inspired by Navier–Stokes equations, moving beyond diffusion-based message passing. This convection-aware design mitigates oversmoothing and enables deeper, more expressive graph models for learning on graphs.
Gaussian Sheaf Neural Networks model node features as Gaussian distributions and use sheaf-theoretic constructions to preserve the geometric and algebraic structure of means and covariances during message passing. This respects the underlying distributional geometry better than naive concatenation.
Is Fixing Schema Graphs Necessary? Full-Resolution Graph Structure Learning for Relational Deep Learning introduces FROG, a method to learn graph structures end-to-end rather than relying on fixed schemas. This full-resolution approach aims to preserve relational semantics directly from data.
EvoStruct bridges evolutionary priors and structural priors for antibody CDR design by adapting protein language models and incorporating substitution patterns from evolutionary databases. This helps counter vocabulary collapse and enables more diverse, functionally relevant design.
GraphRAG benchmarks local LLM-based retrieval augmented generation for healthcare EHR schema retrieval on consumer hardware. The study analyzes reliability, latency, and privacy implications of deploying LLMs locally, providing practical guidance for privacy-preserving retrieval.
Velocityformer is an equivariant graph transformer designed to match the broken symmetry of observational cosmology data for velocity reconstruction. By respecting the data’s symmetry, it improves the accuracy of reconstructed galaxy velocities from spectroscopic surveys, aiding kSZ analyses.
CGSTAE proposes a causal graph spatial-temporal autoencoder for industrial process monitoring, combining a spatial self-attention-based correlation graph learner with a GCLSTM encoder–decoder. The model yields interpretable, dynamic monitoring of processes with a causal perspective.
Federated Learning of Nonlinear Temporal Dynamics with Graph Attention-based Cross-Client Interpretability enables decentralized learning across heterogeneous subsystems. Graph attention provides a way to share interpretable cross-client signals without exposing raw data.
Graph Neural Network based Hierarchy-Aware Embeddings of Knowledge Graphs uses ontology-derived semantic losses to produce hierarchy-aware KG embeddings. The approach is applied to yeast phenotype prediction and also explores box embeddings for KG revisions.
S2Aligner provides pair-efficient and transferable pre-training for sparse text-attributed graphs, enabling robust graph-text alignment when textual anchors are missing or noisy. It reduces reliance on strong supervision while improving cross-domain transfer.
Improved convergence rate of kNN graph Laplacians with differentiable self-tuned affinity defines a learnable affinity that adapts to local densities, leading to faster and more reliable convergence of graph-based Laplacians.
Do Better Volatility Forecasts Lead to Better Portfolios? Evidence from Graph Neural Networks compares GNN-based volatility forecasts to traditional models and examines whether lower forecast error translates into superior portfolio performance.
Discoverable Agent Knowledge proposes a formal framework for agentic knowledge graphs and affordances, enabling coherent discovery, composition, and invocation of KG-enabled capabilities across heterogeneous agents.
Conflict-Resilient Multi-Agent Reasoning via Signed Graph Modeling introduces signed graphs to explicitly model supportive and conflicting inter-agent relations, improving robustness to conflicts in multi-agent reasoning.
BLINKG provides a benchmark for evaluating LLM-integrated knowledge graph generation, focusing on how well input schemas align to ontology terms and how effectively LLMs can generate coherent KGs.
Projecting Latent RL Actions proposes a method to project latent actions to improve generalization and scalability in graph-constrained combinatorial optimization using reinforcement learning and graph neural networks.
Query-Conditioned Graph Retrieval for Contextualized LLM Reasoning in Personalized Wearable Data introduces Wearable As Graph (WAG), a graph-based retrieval framework that enables query-adaptive reasoning over personalized wearable metrics for improved LLM-based inference.
STAR introduces semantic tuning and tail-adaptive retrieval to GraphRAG systems, addressing sparse semantic information in graphs. It mitigates semantic shortcut bias and long-tail path bias, enabling more accurate semantic modeling and improved GraphRAG performance for multi-hop reasoning.
Agentic GraphRAG presents a collaborative agentic framework for expert analysis of commercial registry data. It builds a Neo4j knowledge graph through a three phase pipeline to integrate structured records with unstructured legal text, enabling more effective multi hop, temporal, and entity centric investigations.
Query-Aware Flow Diffusion RAG proposes QAFD-RAG, a diffusion based method for graph RAG with theoretical guarantees on subgraph quality and relevance. By conditioning the diffusion process on the query, it adapts exploration to the user's intent rather than using static neighborhoods.
Automated Big Data Quality Assessment using Knowledge Graph Embeddings proposes a knowledge based method to predict missing edges linking a dataset context to quality rules in a knowledge graph. Embeddings enable context aware quality assessment across data characteristics and required operations.
Graph-Driven Cross-Industry Real-Time Monitoring Framework for Anti-Money Laundering Detection in Converged Mobility-Energy Supply Chain Networks
SCAFDS: Edge-Feature Graph Attention for Interbank Fraud Detection with Attribution-Grounded SAR Generation
HypergraphFormer: Learning Hypergraphs from LLMs for Editable Floor Plan Generation
GOAL: Graph-based Objective-Aligned Diffusion Solvers for Dynamic Multi-Objective Optimization
EgoCoT-Bench: Benchmarking Grounded and Verifiable Operation-Centric Chain of Thought Reasoning for MLLMs
ContextRAG: Extraction-Free Hierarchical Graph Construction for Retrieval-Augmented Generation
TERGAD: Structure-Aware Text-Enhanced Representations for Graph Anomaly Detection
Network-Aware Bilinear Tokenization for Brain Functional Connectivity Representation Learning
DiagEval: Trajectory-Conditioned Diagnosis for Reliable Software Evaluation with GUI Agents
DeTox-Fed: Detecting Toxic Conversations in the Fediverse with Federated Graph Neural Networks