Showing 30 papers for 2026-07-14
This survey presents a two-level taxonomy for GNN-based KG technologies: a knowledge graph technologies pipeline and a GNN-based perspective. It analyzes how KG components map to GNN methodologies and aims to systematize methodologies across the KG technology stack to guide future research.
The paper studies distance-preserving embeddings in inhomogeneous random graphs using landmark-based embeddings to approximate shortest-path distances. It provides theoretical distortion bounds that better reflect typical networks rather than worst-case guarantees and validates the practicality of the approach with experiments.
GAE (Graph-Augmented Evolution) couples evolutionary program search with reinforcement optimization guided by large language models and relational GNNs. It addresses blind parent selection, sparse rewards, and static mutation operators by tightly coupling search with graph-based guidance and learning.
SALT-GNN tackles dense neighborhoods in AML graphs with statistics-aware attention, mitigating signal dilution caused by high-activity recipient accounts. The approach improves detection performance while reducing investigation costs.
The work provides a closed-form theoretical characterization of GNN calibration under distribution shift, showing how calibration depends on specific distributional changes. This yields fundamental insights into when and how GNNs remain reliable under shift.
Graph Neural Networks for RFID-Based Spatial Geometry Inference develops GNN-based methods to infer higher-order spatial relationships in indoor environments from RFID signals. The approach improves robustness to noisy propagation and enhances spatial understanding for Spatial AI systems.
A Novel Graph Fraud Detector via Grouped Attribute Completion and Confidence-Aware Contrastive Learning (GFD-GC) tackles missing attributes and severe class imbalance in graph fraud detection. It combines grouped attribute completion with confidence-aware contrastive learning to improve detection.
Learning Subgroup Relations Using Siamese Graph Neural Networks represents finite groups by Cayley graphs and learns embeddings with a Siamese GNN branch, augmented by algebraic features to predict subgroup relations.
Surprisingly Simple and Effective Multi-Domain Graph Foundation Model through Graph-to-Table Alignment shows that aligning graph representations with tabular features can yield strong cross-domain transfer without heavy text attributes, forming a simple yet effective GFM.
Structure-Feature Aligned Graph Learning via Alternating Constrained Optimization introduces a two-view framework that decouples structure and features by using an independent anchor network to learn a structure-free feature prior, then aligns it to structure-conditioned GNN embeddings via alternating optimization.
A multi-scale feature enhanced graph neural network for fluid dynamics prediction in complex geometries develops multi-scale features to improve GNN-based predictions on large-scale meshes common in fluid dynamics, achieving higher accuracy and efficiency.
MVMGNN; Multi-View Masked Graph Neural Network for Alzheimer's Disease Diagnosis using Structural MRI proposes a multi-view, masked GNN approach to structure MRI data for AD diagnosis, improving robustness and accuracy.
Inverse-IMPRESSION presents a graph-based platform for molecular structure elucidation from experimental NMR properties, built on an inverted Graph Transformer (IMPRESSION-G2). It includes a one-shot bond predictor, a structure-correction stage, and a noise-augmented multi-shot ensemble.
Hyper-modal Imputation Diffusion Embedding with Dual-Distillation for Federated Multimodal Knowledge Graph Completion (MMFeD3-HidE) enables federated MKG completion by securely sharing embeddings via diffusion and dual distillation, improving missing link prediction across clients.
Graph Optimization Foundation Model tokenizes graph problems under a language-model paradigm, enabling a foundation-model-style pretraining and transfer for graph-structured operations research problems.
Local Message-Passing for Discrete Graph Generation introduces GenGNN, a modular message-passing backbone for discrete graph generation without global attention, enabling edge-field persistence and latent refinement; diffusion-based GenGNN variants achieve high validity.
Graph-Based Bayesian Optimization for Quantum Circuit Architecture Search uses a GNN surrogate to model circuits as graphs and employs expected improvement with uncertainty estimates (via Monte Carlo dropout) to guide circuit mutations and selections, evaluated on hybrid quantum-classical platforms.
An Approximate Graph Elicits Detonation Lattice introduces a training-free graph-based method to precisely segment and measure detonation cells from 3D pressure traces, improving upon manual 2D edge-detection methods with synthetic data experiments.
UNIT: Unleash Large Language Models Potential for Graph Continual Learning investigates graph continual learning in streaming settings, addressing semantic-structural separation and imbalanced knowledge transfer to improve knowledge retention and transfer over time.
GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention proposes an entity-side message function with gated rotary attention to encode temporal information, enabling better inductive transfer on temporal knowledge graphs with disjoint vocabularies.
KGCQual introduces an interpretable intrinsic metric to assess Knowledge Graph quality from text. It measures how closely the automatically extracted KG approximates an ideal reference, addressing spurious or incomplete triples produced by pipelines. The metric provides insight into structural and semantic fidelity beyond task-specific evaluations.
This paper compares GAT and BERT for human-like playtesting in puzzle games. It discusses how predictive models trained on diverse player data can simulate player behavior but highlights limitations in capturing the full range of strategies and the need for feature engineering and architectural adaptation when new game mechanics appear. The study offers comparative insights and suggests directions to better model varied player strategies.
The authors model motor cortex dynamics during a wrist-extension BCI task as a port-Hamiltonian system, combining a conservative interconnection with a dissipative port driven by a GNN surrogate. A metriplectic integrator evolves the state, and a fluctuation-dissipation-consistent noise channel yields stochastic trajectories at body temperature. On real EEG data, the approach achieves competitive error metrics and demonstrates scale-consistent dynamics.
PromptGraph is a graph-guided prompt sanitization approach for privacy-preserving LLM inference. It models pairwise relationships among spans to balance privacy and utility rather than sanitizing spans independently. By constructing a prompt graph, PromptGraph achieves more effective privacy protection while preserving task performance.
RAGU introduces a multi-stage GraphRAG engine coupled with a compact domain-adapted LLM. It separates extraction from consolidation through typed two-stage extraction, DBSCAN-based deduplication, LLM summarization, and Leiden community detection. The design emphasizes a compact extractor tailored to the domain to improve retrieval quality and efficiency.
MG^2-RAG proposes a Multi-Granularity Graph for multimodal retrieval-augmented generation. It addresses limitations of flat vector retrieval and translation-to-text pipelines by jointly modeling cross-modal dependencies with a lightweight, multi-granularity graph. The approach reduces hallucinations and improves multimodal reasoning in RAG.
RankGraph-2 presents lifecycle co-design for billion-node graph learning in recommendations. It treats graph construction, representation learning, and real-time serving as coupled stages, deploying a co-trained cluster index to avoid online KNN. The approach tightens the loop between training and serving for scalable, high-quality retrieval.
Conceptual Networks for Cross-Linguistic Idiomatic Expressions uses an interpretable graph framework to map idiomatic meanings across eight languages. Each expression is annotated with binary conceptual features, and pairwise Jaccard similarities define a weighted graph. Community detection reveals idioms cluster by conceptual schema rather than by language.
NGM-RAG introduces Neural Graph Matching-based Retrieval-Augmented Generation. By leveraging graph-structured matching, it enhances retrieval for complex, multi-hop questions beyond text-based methods. The approach demonstrates improved accuracy on challenging queries by exploiting graph relationships.
FAIR GraphRAG proposes a retrieval-augmented generation framework that aligns with FAIR data principles for semantic data analysis. It argues that existing RAG approaches lack structured FAIRification of the underlying knowledge graphs and outlines methods to enhance findability, accessibility, interoperability, and reusability within graph-enabled analyses.