Showing 33 papers for 2026-05-19
This paper proposes a Bi-Level Chaotic Fusion based Graph Convolutional Network to predict prediction intervals for stock prices, addressing uncertainty in financial forecasting. It leverages chaotic fusion at two levels to produce upper and lower bounds, providing uncertainty quantification to support risk-aware decisions.
We introduce Hypergraph Pattern Machine, a compositional tokenization method for higher-order interactions on hypergraphs. It captures interaction compositionality (compositional, emergent, or inhibitory) in higher-order relations and supports safe simplification or joint reasoning in domains like polypharmacy.
This work studies Graph Structure Learning (GSL) and its tendency to explode edge counts as it seeks more informative structures. It proposes mechanisms to keep the learned graph informative while avoiding excessive edge growth, improving efficiency without sacrificing performance.
We introduce a tensor-channel equivariant GNN for direct prediction of molecular polarizability tensors. The model augments hidden representations with symmetric rank-2 tensor channels, aligned with isotropic and anisotropic components, and propagates tensor structure through message passing using geometrically motivated bases. This yields predictions that are better aligned with physical symmetries and improves target accuracy.
We present UNR-Explainer, a method to generate counterfactual explanations for unsupervised node representation learning. It identifies the most influential subgraphs whose perturbation markedly changes the k-nearest neighbors of a node in the learned embedding space.
We address missing structured data in spatiotemporal sensor networks by proposing multi-scale hypergraph Laplacians with neural refinement to capture higher-order relations among sensor groups. This approach improves imputation by leveraging structured absences and higher-order constraints beyond pairwise graphs.
We argue that negative edges carry informative signal and develop a contrastive message-passing framework that uses both positive and negative edges. Under regimes with low label rates, high homophily, and dense graphs, incorporating negative edges yields significant information gain and improves representations.
We introduce Graph Hierarchical Recurrence (GHR), a framework that combines hierarchical recurrence with GNNs/Graph Transformers to capture long-range dependencies in graphs. This improves generalization on tasks requiring distant node correlations.
We propose S2Aligner, a pre-training framework for sparse text-attributed graphs that aligns graph structure and node texts using weak semantic evidence. It is pair-efficient and transferable across domains, addressing missing, noisy, or uneven textual anchors to enable robust structure-text alignment.
LARGER introduces a lexically anchored repository graph approach to exploration and retrieval, enabling better navigation beyond plain lexical search by recovering structural relations like imports and call chains. The method builds a graph-based retrieval system that leverages structural dependencies to locate relevant files and symbols.
This work studies backdoor defenses beyond subgraph triggers, focusing on feature-based backdoors. It analyzes how homophily influences backdoor vulnerability and proposes defense strategies from a feature-centric perspective to improve robustness.
We propose a two-stage Filter-then-Verify framework combining an inductive GNN for structural anomaly detection with ModernBERT for content verification. The GNN flags anomalous sender-receiver patterns, while BERT analyzes message semantics to reduce false positives; experiments on Enron with synthetic campaigns show improved detection.
SynC proposes a synergistic framework that jointly improves structure augmentation and representation learning to enhance deep graph clustering. It emphasizes mutual reinforcement—the more homogeneous the graph, the better representations, and the better structure augmentation becomes, improving performance on low-homophily graphs.
We generalize graph isomorphism to temporal graphs via consistent event graph isomorphism, capturing the causal topology induced by the arrow of time. This provides a formal lens to analyze the expressive power of temporal graph neural networks and their ability to distinguish temporal patterns.
We address missing modalities in multimodal cancer classification with a missing-modality-aware GNN that robustly fuses available modalities without requiring imputation or fixed fusion. The method scales to many modality combinations and improves classification accuracy across datasets.
We propose to utilize historical activations from earlier GNN layers to improve pooling and downstream performance, rather than relying solely on the last layer.
The paper challenges current GNNs' handling of missing node features, showing that high sparsity can mask true robustness and that existing evaluations can be misleading under missingness. They propose robust evaluation and a practical solution to improve resilience to missing features.
The authors comprehensively analyze full-graph and mini-batch GNN training, introducing batch size and fan-out as key factors. They provide insights and guidelines on when to prefer full-graph versus mini-batch training based on convergence, generalization, and computational costs.
They study domain adaptation for physics-informed self-attention GNNs in AC power flow prediction across voltage regimes, focusing on parameter-efficient methods to transfer knowledge with limited retraining. The approach balances stability and plasticity for cross-regime performance.
Flowette is a continuous flow-matching framework for graph generation that uses a graph neural network-based transformer to learn a velocity field over graph representations. It promotes topology-aware alignment via optimal transport and introduces graphettes as priors to encode domain-driven subgraph motifs, enabling generation of graphs with recurring motifs.
STAG-CN introduces a spatio-temporal apiary graph convolutional network for disease onset prediction across apiaries. It builds a dual adjacency graph that combines physical co-location and climate-driven sensor correlations across hive sessions, and processes multivariate IoT sensor data with graph convolutional layers. By modeling inter-hive disease pathways, STAG-CN aims to enable earlier and more reliable outbreak warnings.
We present Graph Neural ODE Digital Twins for control-oriented forecasting of reactor thermal-hydraulic states under partial observability. The approach uses a physics-informed message-passing Graph Neural Network coupled with Neural Ordinary Differential Equations (GNN-ODE) to achieve high predictive fidelity, millisecond-scale inference, and robustness when sensors are unavailable. It models the plant-wide dynamics with partial observability to support real-time control.
We propose Zero-Shot Scalable Resilience in UAV Swarms: a decentralized imitation-learning framework with physics-informed graph interactions. The method avoids reliance on centralized topology by using decentralized imitation and graph-based interactions that encode physics-informed constraints. It aims to enable scalable, zero-shot resilience across varying swarm sizes and damage scenarios.
SuReNav introduces a superpixel graph-based constraint-relaxation approach for navigation in over-constrained environments. It relaxes hard constraint regions into softer costs via a superpixel graph to seek a best-effort path that avoids restricted areas while preferring low-risk regions. This improves generalization beyond predefined area costs and enables safer navigation in semi-static scenes.
Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks explains combining LMs and GNNs to form foundation models for relational databases. It argues that traditional approaches flatten data and lose relational context, and that relational entity graphs enable richer, task-agnostic learning. The framework aims to generalize relational learning across tasks and databases.
Attention Dispersion in Dynamic Graph Transformers: Diagnosis and a Transferable Fix investigates why CTDG transformers under temporal distribution shift. It identifies attention dispersion as a shared failure mode and shows that a small set of critical nodes drive predictions. The paper then proposes a transferable mitigation to improve robustness across datasets and models.
Graph-Regularized Sparse Autoencoders for LLM Safety Steering introduces GSAE, which learns safety-oriented directions by smoothing decoder vectors over a neuron co-activation graph. This dictionary-learning approach encodes distributed activation structure tied to safe behaviors, improving steering decisions for large language models.
Stock Market Prediction Using Node Transformer Architecture Integrated with BERT Sentiment Analysis presents a framework that combines a node-transformer model with BERT-based sentiment signals to forecast stock prices. The approach seeks to capture cross-sectional dependencies and market sentiment to improve predictive performance in noisy, non-stationary markets.
Region-Grounded Report Generation for 3D Medical Imaging introduces VietPET-RoI, a large-scale 3D PET/CT dataset with fine-grained region-of-interest annotations for low-resource languages, and a graph-enhanced framework for region-grounded report generation. The approach grounds reports to anatomical RoIs to improve accuracy and interpretability.
ShadowMerge reveals a poisoning attack on graph-based agent memory by injecting relation-channel conflicts. An attacker can insert crafted relations into the graph memory to skew retrieval and influence agent behavior, highlighting a new vulnerability in graph-based memory beyond flat textual attacks.
Representing Higher-Order Networks: A Survey of Graph-Based Frameworks surveys higher-order graph formalisms that extend beyond pairwise interactions, including multiway, hierarchical, temporal, multilayer, and recursive structures, with mathematical foundations. It provides a comprehensive taxonomy and toolkit for modeling complex systems.
RCTEA: Richness-guided Co-training for Temporal Entity Alignment proposes a framework that jointly leverages structural and temporal features and information richness to improve alignment of entities across temporal knowledge graphs. It emphasizes orthogonal effects and richer message passing for TEA.
Agentic Chunking and Bayesian De-chunking of AI Generated Fuzzy Cognitive Maps presents an approach to automatically generate feedback causal fuzzy cognitive maps (FCMs) from text by teaching LLM agents to chunk text into overlapping pieces and then convexly combine them. The resulting cyclic FCM enables Bayesian inference to model dynamic relationships such as the Thucydides Trap.