Showing 13 papers for 2026-08-03
We propose subject-adaptive MPP-GNN for fMRI-based Alzheimer's disease classification. Instead of using a fixed set of brain modules for all subjects, the model learns per-subject modular decompositions and uses these modules to guide the learned connectivity, improving classification accuracy and interpretability.
We introduce MMFGU, a framework for multimodal federated graph unlearning. It supports fine-grained deletion requests, such as removing entities, interactions, or entire modalities, while preserving model performance and privacy. The paper develops efficient unlearning procedures and analyzes the privacy-utility trade-offs.
We develop a value-based reinforcement-learning framework for dynamic matching on finite, evolving weighted graphs. The approach handles arrivals, transitions, edge realizations, and exogenous exits to optimize matching decisions over time. We prove an event-time reduction showing that it is optimal to act immediately after each exogenous event and then wait for the next event.
We study how Graph Foundation Models can transfer knowledge across unseen domains and propose a method to identify and align transferable propagation knowledge. By focusing on propagation units rather than fixed tokens, the approach aims to generalize GFMs to new graph domains.
We systematically benchmark spatio-temporal graph neural networks for fault location under increasing DER penetration. The study analyzes robustness, generalization, and performance trends as intermittent DER generation and bidirectional power flows alter fault signatures. The results reveal strengths and limitations of STGNNs for distribution fault location scenarios.
We introduce Cross-Resolution Semantic Learning for Graph Domain Adaptation. The core idea is to model propagation resolutions (neighborhood ranges) and a semantic-resolution shift that routes class-discriminative knowledge to appropriate target resolutions, enabling better cross-domain transfer of graph knowledge.
We propose TFGformer for multivariate time-series forecasting by combining time-frequency graph learning with covariate fusion. The model captures inter-series relationships via graphs on the time-frequency domain and integrates external covariates, addressing nonstationary patterns. It enables improved long-horizon forecasting for large-scale IoT sensor data.
We investigate ordered-to-disordered transfer learning with graph neural networks for predicting formation energy and HOMO-LUMO gaps in high-entropy perovskite oxides. By transferring knowledge from ordered configurations to disordered ones, the approach achieves accurate property predictions with limited data and data-efficient training.
We propose DiRe, a force-directed dimensionality reduction framework that emphasizes global structure and homological features while remaining scalable. It combines an initial embedding with a graph-based layout optimization and evaluates representations with distortion, context preservation, and persistent homology metrics, offering a complementary alternative to UMAP/t-SNE.
We present a GNN-based multi-agent control scheme for mitigating traffic shockwaves in sparse VANETs. The decentralized approach avoids reliance on global traffic state information and coordinates vehicles through neighbor-aware graph reasoning, reducing stop-and-go waves and improving flow efficiency.
We design an ontology-guided, deduplication-aware extraction layer to construct knowledge graphs from heterogeneous documents. The layer reconciles inconsistent vocabularies, name variants, and duplicated relations, producing validated graphs aligned to a formal ontology, and supports production-scale deployment.
MolSight is a graph-aware vision-language model for unified chemical image understanding. By integrating graph representations of molecular structures with vision-language reasoning, it improves structural alignment and topological modeling for tasks such as property prediction and image interpretation.
We propose MERIT, an efficient in-place deletion method for dynamic graph-based approximate nearest neighbor indexes. It supports removing obsolete vectors without propagating stale edges or performing expensive graph-wide maintenance, preserving online service throughput.