Showing 19 papers for 2026-06-25
This paper analyzes whether Holographic Reduced Representations (HRR) and phase-only Fourier HRR can support zero-shot compositional reasoning in knowledge graphs. By comparing real-valued HRR and FHRR on multi-hop queries, the authors inspect the binding/unbinding mechanics, invertibility, and where these methods fail, offering mechanistic explanations for their limitations.
We propose DCQ-GNN, a spectral GNN that uses a compact two-tap bank of convex–concave quadratic filters to achieve sharper spectral selectivity. By keeping the filter order at two and leveraging complementary curvature, DCQ-GNN improves frequency localization and attenuation outside the passband, addressing challenges of high-order filters.
The paper introduces Auto-Configured Explainable GNNs for multi-site pollution prediction. It builds a dynamic graph by constructing inter-class edges from a confusion matrix produced in a supervised stage, capturing relationships between classes. A hybrid loss combining energy distance with Huber loss is used to mitigate vanishing gradients and improve predictive performance across sites.
GCT-MARL proposes a transfer framework for cooperative multi-agent RL. It builds on a multi-view graph contrastive backbone and adds a per-view adaptively weighted alignment loss and a two-phase training protocol designed for transfers across populations with differing sizes and compositions, achieving improved sample efficiency.
CIPE (Communicability-Inspired Positional Encoding) introduces a new positional encoding for graphs based on communicability, aggregating contributions from paths of all lengths. The encoding aims to be compatible with attention by shaping a geometry in which node similarities reflect meaningful graph structure, improving Transformer-like processing on non-Euclidean graphs.
The paper conducts the first comprehensive gradient leakage attack evaluation on GNNs in circuit-design and hardware-security tasks. It tests SOTA GNNs (GraphSAGE, GCN, GIN, GAT) on standard netlists and demonstrates fundamental vulnerabilities to GLA, highlighting privacy and security risks in critical applications.
The authors propose hierarchical graph learning for calendar spread strategies in commodity futures. They represent markets hierarchically with assets and contracts connected by cross-level edges reflecting correlations, and develop a hierarchical GNN approach to capture both intra- and inter-level relationships for better CS strategy.
The study evaluates whether risk stratification for pulmonary embolism can be achieved from medical records and CTPA biomarker data when blood tests are missing. It benchmarks models that combine clinical records with cardiac biomarkers extracted from CTPA, comparing performance to guidelines-based approaches.
This work studies instance hardness in combinatorial auctions by training GNNs to predict when a given WDP instance will be hard for greedy allocation. Rather than replacing solvers, it aims to enable instance-dependent algorithm selection, improving performance by routing easy instances to fast heuristics and hard ones to more powerful methods.
The authors propose fast condition-number estimation for sparse matrices using graph neural networks. They design graph features with linear complexity O(nnz+n) and present two schemes to estimate the condition number by decomposing it into components, balancing accuracy and efficiency for practical deployment.
SDE-HGNN introduces an SDE-driven spatio-temporal hypergraph neural network for irregular longitudinal fMRI connectome modeling in Alzheimer's disease. It uses an SDE-based reconstruction module to recover continuous latent trajectories from irregular observations, enabling robust modeling of disease progression.
Ramanujan Graph Rewiring with Non Negative Resistance Curvature
KIGNet: Physics-Motivated Multi-Graph Representation Learning for Explainable Jet Tagging
A Hybrid TGN-SEAL Model for Dynamic Graph Link Prediction
Position Spaces and Graphs
Fuzzy Quantification over OWL Ontologies and Knowledge Graphs
Physics Question Scene Graph: Fine-grained Evaluation of Physical Plausibility in Text-to-Video Generation
Is GraphRAG Needed? From Basic RAG to Graph-/Agentic Solutions with Context Optimization
CausalRAG2: Hierarchical Causal Knowledge Graph Design for RAG