Showing 42 papers for 2026-04-14
K-STEMIT proposes a knowledge-informed spatio-temporal multi-branch graph neural network to estimate subsurface stratigraphy thickness from radar data. By integrating domain knowledge with graph-based processing, it aims to be robust to radar speckle noise and acquisition artifacts while delivering accurate thickness estimates across space and time.
Graph-RHO introduces a critical-path-aware heterogeneous graph network to tackle long-horizon Flexible Job-Shop Scheduling (FJSP). It captures complex graph-structured dependencies and accounts for asymmetric costs of prediction errors within a rolling-horizon optimization framework, improving solution quality and planning speed.
A Temporally Augmented Graph Attention Network (EEG-tGAT) adds explicit temporal modeling to graph attention, enabling affordance classification from interaction sequences. It integrates temporal context into node attention to capture evolving relational cues that improve classification accuracy.
Virtual Smart Metering in District Heating Networks proposes using heterogeneous spatial-temporal graph neural networks to augment observability when sensors are sparse or faulty. The approach enables virtual sensing, improving reliability, predictive control, and fault detection in thermal networks.
A diffusion-contrastive graph neural network with virtual nodes extends wind nowcasting into unobserved regions without installing new sensors. It uses self-supervised diffusion and contrastive learning to propagate information across the network and improve forecasts.
Topology-aware PAC-Bayesian generalization analysis for graph neural networks provides data-dependent, topology-sensitive generalization bounds for graph classification. The work elucidates how graph structure interacts with model parameters to affect generalization.
DIB-OD preserves an invariant core for robust heterogeneous graph adaptation using a decoupled information bottleneck and online distillation. The framework disentangles task-relevant invariant knowledge from domain-specific noise to reduce negative transfer and forgetting.
Hypergraph Neural Diffusion (HND) unifies nonlinear diffusion equations with neural message passing on hypergraphs. Grounded in a continuous-time diffusion model, it addresses shallow propagation, oversmoothing, and adaptability to complex hypergraph structures.
Unified Graph Prompt Learning via Low-Rank Graph Message Prompting proposes a single prompter that targets all graph components (nodes, edges, and weights) with low-rank prompting to efficiently adapt pre-trained GNNs. It unifies the prompting space for graph data.
Sheaf Diffusion with Adaptive Local Structure for Spatio-Temporal Forecasting reformulates forecasting as information flow over locally structured spaces using ST-Sheaf GNNs. It encodes topology into diffusion to capture higher-order interactions under local heterogeneity.
Learning How Much to Think: Difficulty-Aware Dynamic MoEs for Graph Node Classification introduces D2MoE, a node-wise dynamic mixture-of-experts that assigns different numbers of experts based on node difficulty. This yields more efficient and accurate classification.
Generating Multiple-Choice Knowledge Questions with Interpretable Difficulty Estimation using Knowledge Graphs and Large Language Models builds MCQs by constructing a knowledge graph from documents with an LLM and then selecting a key node and related edges to generate questions. Difficulty estimates are interpretable.
Panoptic Pairwise Distortion Graph introduces Distortion Graphs to represent paired images as a structured region-level topology for comparing degradations. It formalizes a panoptic inter-image graph-based framework for distortion assessment.
Frugal Knowledge Graph Construction with Local LLMs demonstrates a zero-shot pipeline using local language models, with self-consistency and artificial crowds insights, achieving competitive F1 on DocRED/HotpotQA-like data in a reproducible framework.
SIGMA: An Efficient Heterophilous Graph Neural Network with Fast Global Aggregation proposes a scalable GNN that handles heterophily efficiently via a global aggregation scheme, enabling fast processing on large-scale graphs.
Adversarial Robustness of Graph Transformers designs adaptive gradient-based attacks to study the robustness of Graph Transformer models, providing principled attack strategies and evaluating across multiple GT architectures.
ExPath: Targeted Pathway Inference for Biological Knowledge Bases via Graph Learning and Explanation frames a subgraph inference task that integrates experimental data to classify bio-networks and explain which links contribute most to the decision.
Soft Graph Transformer for MIMO Detection introduces the Soft Graph Transformer (SGT), a soft-input-soft-output detector that leverages the graph-structured factor graph and prior information to improve MIMO detection with feasible complexity.
Discrete Bayesian Sample Inference for Graph Generation presents GraphBSI, a one-shot graph generator based on Bayesian inference that iteratively refines a belief over graphs in a continuous distribution space rather than evolving discrete graphs.
AdvSynGNN combines structure-adaptive graph neural networks with adversarial synthesis and self-corrective propagation. It uses multi-resolution structural synthesis and contrastive objectives in a transformer backbone to improve robustness under structural noise and heterophily.
We introduce GIST, Gauge-Invariant Spectral Transformers, to build scalable graph neural operators without cubic eigendecomposition. GIST preserves gauge invariance in spectral approximations to avoid solver artifacts that hurt generalization when graphs are discretized differently. This yields a scalable, generalizable spectral transformer for graphs and meshes across varying discretizations.
PACIFIER reframes polarization moderation as a graph-based sequential planning problem and proposes a graph reinforcement learning framework. It moves beyond Friedkin–Johnson model-based linear steady-state interventions, enabling scalable, data-driven pacing of interventions to curb opinion polarization. The approach demonstrates improved handling of complex, networked intervention regimes.
SEARL proposes joint optimization of a policy and a tool-graph memory to support self-evolving agents. It tackles resource-constrained settings where outcomes are sparse, leveraging trajectory-based learning and tool usage without relying on large LLMs or multi-agent systems. The framework improves sample efficiency and enables evolving capabilities in agents.
CID-TKG is a collaborative learning framework for temporal knowledge graph reasoning that integrates evolutionary dynamics and historical invariance. By coupling dynamic evolution with invariance properties, it improves reasoning over evolving facts across time. Experiments show better predictive accuracy and robustness over baselines.
EL-DRUIN presents ontological trajectory forecasting using finite semigroup iteration and Lie algebra approximation in geopolitical knowledge graphs. The system models relationships as dynamic patterns inside a finite state space and composes them with semigroup operations, offering long-horizon forecasts beyond pattern-matching LLMs. It aims to produce explainable, ontology-grounded trajectory predictions.
Proposes inductive reasoning for temporal knowledge graphs with emerging entities. It addresses the open-world challenge where new entities appear without historical interactions, showing methods that generalize to unseen nodes. The approach improves forecasting for emerging entities in evolving graphs.
Provides a benchmark for gap and overlap analysis for policy-like documents, focusing on competency questions. Given a scenario, it identifies which documents support it (overlap) and which do not (gap) with defensible justifications. This task-ready evaluation assesses KG readiness to answer real-world questions.
Proposes a scheduler-theoretic framework that replaces the Agent Loop with a structured graph of executable units. At any time, only one unit runs, and the next activation is chosen by LLM inference, reducing hidden dependencies and unbounded recovery loops. The framework improves debugging, reliability, and predictability of LLM-based agents.
Systematically evaluates graph-based retrieval and agentic retrieval for cyber threat intelligence. It shows that knowledge graphs better support reasoning across distributed fragments by modeling relationships among actors, malware, and vulnerabilities. The study highlights when graph-based and agentic retrieval outperform vanilla RAG in CTI tasks.
DreamKG is a KG-augmented conversational system for people experiencing homelessness. It grounds responses in verified, up-to-date Philadelphia service data and uses Neo4j plus structured query understanding to handle location- and time-sensitive queries. This reduces hallucinations and improves reliability when guiding PEH to services.
Do We Still Need GraphRAG benchmarks RAG and GraphRAG in agentic search systems. It experiments with dynamic, multi-round retrieval and sequential decision-making, showing graph-based retrieval improves evidence gathering and reasoning beyond standard RAG.
Investigates structural complexity in normative documents with graph-based approaches, using ETSI standards as a case study. It argues vanilla vector-based retrieval misses latent structural and relational features and demonstrates how graph representations capture hierarchy and cross-references to improve retrieval and understanding.
A dual cross-attention graph learning framework for multimodal MRI-based major depressive disorder detection. The framework fuses structural and functional MRI via bidirectional cross-attention, enabling interactions between modalities to improve detection performance.
Time is Not a Label introduces RoMem, a continuous phase rotation memory module for temporal knowledge graphs and agent memory. It separates persistent facts from evolving ones, enabling continual, time-aware memory without overwriting or excessive recomputation. RoMem provides structured, scalable memory for long-lived agents.
Think Parallax argues multi-hop reasoning is inherently multi-view. It proposes a KG-RAG approach that distributes hops across distinct attention-head views, enabling hop-aligned relay pathways. This multi-view retrieval improves accuracy on multi-hop tasks over single-view KG-RAG baselines.
AtlasKV augments LLMs with billion-scale knowledge graphs in 20GB VRAM using a parametric integration method. It reduces retrieval latency and memory footprint while delivering vast external knowledge to LLMs, enabling scalable knowledge augmentation under tight hardware constraints.
M3KG-RAG extends RAG to multi-hop multimodal knowledge graphs. It addresses limited modality coverage and weak multi-hop connectivity in existing MMKGs by designing retrieval that leverages multimodal graph structure, yielding more relevant and comprehensive responses in audio-visual domains.
GRAPHIA is a general LLM-based social simulation framework that uses social-graph data as supervision. It trains agents via reinforcement learning with GNN-based structural rewards to predict whom to interact with, producing more realistic social dynamics for simulations.
Multi-Faceted Continual KG Embedding (CKGE) learns evolving entity semantics with a multi-faceted representation to mitigate forgetting. By modeling entities with multiple semantic facets that adapt over time, the approach enables semantic-aware link prediction in continual settings.
EA-Agent introduces a structured multi-step reasoning agent for entity alignment. By using explicit, stepwise reasoning and evidence-driven prompts, it improves alignment accuracy under noisy or sparsely supervised conditions, addressing limitations of end-to-end LLM approaches.
ARK proposes a fine-tuning framework for Retrieval-Augmented Generation that optimizes the retriever for Answer Alignment in long-context settings. By incorporating a KG-augmented curriculum, the method guides the retriever toward evidence that more directly improves answer quality, addressing the misalignment between query–document similarity and downstream answer generation.
The study builds a concept–object knowledge graph from the astro-ph corpus up to July 2025 by extracting named astronomical objects, resolving them to SIMBAD identifiers, and linking them to annotated scientific concepts. It then evaluates whether the historical graph structure can anticipate new concept–object associations before they appear in print, leveraging clustering-derived concepts and inference-time similarity to predict future links.