Showing 19 papers for 2026-05-22
In transductive learning on graphs with a full observed graph but only partially labeled nodes, the training objective is orthogonal to architectural advances. The authors show that unlabeled-node predictions during training contain useful signals that standard supervised losses discard. They propose Graph Transductive Sharpening to leverage these unlabeled predictions, providing additional supervision and improving semi-supervised node classification.
WaveGraphNet presents a physics-consistent, coupled inverse-forward graph-learning framework for damage localization using guided waves in composite plates. By constraining the learning with physical relations, it reduces overfitting to training locations and improves localization for unseen regions. The approach aims to generalize damage localization across a broader set of scenarios with sparse sensor networks.
TriForces introduces a model-agnostic three-stream framework that separates composition and structure information in atomistic GNNs to improve transferability across chemistries with limited data. The design preserves physical semantics and can be combined with existing MLIPs to enhance data efficiency. This yields more transferable representations for diverse chemical spaces.
Graph Navier Stokes Networks (GNSN) go beyond diffusion-based message passing by incorporating convection terms inspired by the Navier–Stokes equations. This helps mitigate the oversmoothing problem in deep GNNs and enables more expressive dynamics modeling on graphs. The architecture aims to capture richer spatial-temporal interactions on graph-structured data.
Gaussian Sheaf Neural Networks extend GNNs to distribution-valued node features, where each node is a Gaussian characterized by mean and covariance. By preserving the geometric and algebraic structure of Gaussians in message passing, the method enables uncertainty-aware and structurally faithful inference. This advances learning when features are probabilistic rather than deterministic.
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 without fixed schemas. This preserves relational semantics while adapting topology to data, enabling more flexible relational reasoning on databases.
EvoStruct bridges evolutionary and structural priors for antibody CDR design by adapting protein language models, addressing vocabulary collapse in equivariant GNNs and embedding evolutionary substitution patterns into design. The approach yields more diverse, functionally relevant CDR designs by leveraging evolutionary information.
GraphRAG for EHR schema retrieval is evaluated on consumer hardware with locally deployed LLMs, examining feasibility, latency, cost, and privacy trade-offs in healthcare contexts. The study provides guidance for deploying retrieval-augmented generation without cloud dependencies.
Velocityformer is an equivariant graph transformer designed to match the broken symmetry of cosmological data to reconstruct galaxy velocities from spectroscopic surveys, improving the correlation with true velocities and tightening kSZ inferences.
Graph Autoencoder for Process Monitoring proposes CGSTAE, combining a correlation-graph self-attention module with a spatial-temporal encoder-decoder using GCLSTM to capture dynamic relationships and causality for reliable process monitoring.
Federated Learning of Nonlinear Temporal Dynamics with Graph Attention-based Cross-Client Interpretability presents a federated framework that uses graph attention to learn nonlinear temporal dynamics across heterogeneous clients while preserving privacy, and provides interpretable cross-client attention analyses.
Graph Neural Network based Hierarchy-Aware Embeddings of Knowledge Graphs uses ontology-derived semantic losses to produce hierarchy-aware embeddings, improving yeast phenotype prediction and enabling box embeddings for evaluating KG revisions.
S2Aligner offers pair-efficient, transferable pre-training for sparse text-attributed graphs, addressing missing or noisy textual anchors with robust graph-text alignment under weak supervision to improve cross-domain transfer.
Improved convergence rate of kNN graph Laplacians proposes a differentiable self-tuned affinity for kNN graphs, with adaptive bandwidth based on local density, improving convergence rates of Laplacian-based methods.
Do Better Volatility Forecasts Lead to Better Portfolios? Evidence from Graph Neural Networks compares GNN-based volatility forecasts against AR/LSTM baselines on S&P 500 data, finding that the best forecast accuracy aligns with stronger portfolio performance.
Hypergraph Enterprise Agentic Reasoner over Heterogeneous Business Systems introduces HEAR, an enterprise agentic reasoner built on a stratified hypergraph ontology, with a provenance-aware data interface and a hyperedge layer encoding n-ary rules for grounded, auditable reasoning.
MASFactory presents a graph-centric framework for orchestrating LLM-based multi-agent systems, modeling workflows as computation graphs to enable reuse, integration of diverse contexts, and scalable collaboration among agents.
Evaluation of Pipelines for Data Integration into Knowledge Graphs introduces KGI-Bench to benchmark pipelines that ingest new data into existing KGs, analyzing outputs to compare quality and performance across pipelines.
OSM+: Billion-Level OpenStreetMap Dataset for City-wide Experiments releases OSM+, a billion-scale OpenStreetMap dataset built with distributed cloud computing on 5,000 cores, enabling city-wide graph learning experiments and benchmarking.