Showing 31 papers for 2026-06-18
The paper identifies that multimodal neuroimaging graphs, combining fMRI functional connectivity and DTI structural connectivity, are confounded by demographic factors like age and sex, causing GNNs to rely on spurious shortcuts. It introduces Artemis, an anatomy-resolved intervention framework to remove these confounders and enable causally invariant representations, addressing limitations of prior domain-agnostic causal GNN approaches.
We introduce a K-Hop Gaussian diffusion kernel (KHG) to extend information propagation beyond single-hop neighborhoods. KHG better handles noisy or poorly defined edges and complex local structures by applying a Gaussian-weighted diffusion across k-hops, addressing limitations of existing global diffusion kernels like PPR and Heat Kernel. This leads to improved propagation and performance on graphs with irregular local topology.
Credit card fraud detection faces data imbalance, evolving fraud patterns, and complex relational structures. We propose Time-aware Multi-Relational Guided Graph Neural Network (TMR GGNN), extending encoder-decoder GNNs to model heterogeneous interactions among customers, merchants, devices, and IPs within temporal windows, to capture evolving relational dynamics for improved detection.
Single-cell RNA sequencing clustering is hindered by sparsity and noise as well as intricate intercellular structure. scGTN introduces a deep Siamese Graph Transformer Network to leverage cell-to-cell similarities and intercellular graph structure, yielding more robust clustering.
Predictive process monitoring requires generating structurally feasible full event sequences. Graph Grounded Cross Attention Transformer Neural Network (GGATN) enforces feasibility, temporal order, termination, and attribute consistency through graph-grounded cross-attention, enabling unified generation of complete event sequences beyond single-step predictions.
INDEQS augments Neural Controlled Differential Equations (NCDE) forecasting by incorporating prior directed graph knowledge at selected architectural positions. It separates inner state mixing across graph nodes from cross-graph dynamics, enabling informed temporal forecasting that leverages known topology.
We present Anisotropic Graph Diffusion Network (AGDN) to tackle TSP by exploiting richer graph structure in fully connected graphs. AGDN addresses two core issues: the lack of informative topological priors and the risk of losing connected nodes during diffusion, enabling more effective route learning.
We propose Physics-augmented Koopman-enhanced Graph Convolutional Network (P-K-GCN) for spatiotemporal super-resolution on irregular geometries. By coupling physics constraints with a Koopman operator-enhanced GCN, the model achieves higher fidelity reconstructions of complex spatiotemporal dynamics.
A multimodal ablation study for precipitation nowcasting with Graph Neural Networks assesses the contribution of radar history, MEPS NWP, Netatmo observations, MSG channels, stochastic noise, and CRPS-based ensemble losses. The study clarifies the value of each modality for accurate, timely forecasts.
We introduce a data-driven algebraic multigrid (AMG) smoother using a graph network to predict optimal polynomial coefficients for a sparse pseudo-inverse across diverse grid topologies, accelerating pressure solves in incompressible flows.
We study depth-width tradeoffs in Transformer-based algorithmic graph reasoning, particularly when width can grow linearly. The work analyzes how increased width affects the minimal depth required for solving graph tasks, contributing to the understanding of model efficiency limits.
UST-GNN provides a Unified Spatial–Topological Graph Neural Network framework for urban analytics, integrating spatial autocorrelation and network topology to better predict urban outcomes, demonstrated through urban health prediction.
SME-HGT presents a Heterogeneous Graph Transformer for detecting high-potential SMEs using public data, constructing a heterogeneous graph with company, topic, and agency nodes to predict SBIR Phase II progression.
We model short-term-to-long-term memory transfer in temporal knowledge graphs under partial observability, formulating a neuro-symbolic value-based decision problem. Per-item Q-learning handles variable-sized short-term buffers to decide which observations to retain.
We formalize and mitigate structural distortion in LLM attention when reasoning over graph data. The work analyzes how rotary positional embeddings cause bandwidth-dependent attention decay during graph linearization and proposes mitigation strategies.
We propose fully geometric multi-hop reasoning on knowledge graphs with transitive relations, mapping logical operations to pure geometric transformations. This offers a interpretable geometric framework for multi-hop KG reasoning without heavy neural operators.
A finite-element-inspired bipartite graph learned simulator is developed for manufacturability assessment in large-deformation sheet forming. It replaces costly explicit FE simulations with a bipartite graph model that predicts one-step FE state transitions and rolls out autoregressively.
FlowRAG advances retrieval-augmented generation with FlowRAG: Frequency-aware Multi-Granularity Graph Flow. It mitigates under-retrieval for abstract queries and brittle multi-hop reasoning by enabling frequency-aware, multi-granularity propagation over graphs.
We analyze the structural preservation and logical expressiveness of Graph Neural Networks, linking architectural choices (aggregation, combination, activation) to formal logical capabilities. The work provides a semantic perspective on what GNNs can express in logical terms.
A Unified Framework for Context-Aware and Relation-Aware Graph Retrieval-Augmented Generation integrates context and relational signals in retrieval-augmented generation, overcoming limitations of entity-centric and chunk-centric approaches and enabling emergent knowledge fusion.
This paper proposes a timestamp-aware spatio-temporal graph contrastive learning method for network intrusion detection. By modeling evolving traffic-flow relationships over time, it captures temporal dynamics that conventional GNNs overlook and aims to improve generalization to unseen attacks with less reliance on labeled data.
This study investigates using Graph Neural Networks to support cybersecurity and drone management in a cyber-physical system with intrusions and UAVs. It builds an integrated procedure that lets intrusion detection systems learn from the underlying network structure via GNNs, enabling more informed defense and drone operations.
We propose a trust-aware coordination framework for multi-agent software engineering that uses confidence-calibrated knowledge graphs to maintain traceability and consistency across artifacts. By propagating decision confidence through a shared graph, downstream decisions are better aligned and risks from upstream errors are reduced.
We present Graph Neural Networks for semi-supervised image classification with multi-feature aggregation. The approach combines diverse feature representations from multiple extractors (e.g., CNNs and ViTs) at the graph level to improve label-efficient performance on images.
We show that simply concatenating LLM-derived node features into GNN inputs can harm performance on homophilous benchmarks. Pure concatenation, without joint training, distillation, or prompt-conditioning, can degrade accuracy even where end-to-end LLM pipelines succeed.
We introduce learnable graph patches to address feature heterogeneity without relying on textual information. These patches are the smallest semantic units that enable a universal graph foundation model to transfer across datasets by composing patches into graphs.
We propose enhanced negative sampling to boost Knowledge Graph Foundation Models for zero-shot completion on unseen KGs. By using more informative, type- and relation-aware negative triples instead of random negatives, the training signal becomes stronger and the model generalizes better.
We propose temporal motif-aware graph test-time adaptation for out-of-distribution blockchain anomaly detection. The method leverages evolving transaction motifs to adapt graph representations at inference time, addressing adversarial pattern evolution and varied blockchain semantics.
We introduce GMN4AD, a graph matching network for Alzheimer's disease diagnosis that includes test-time domain adaptation on multi-centered MRI data. By matching brain-region graphs and adapting at test time to site-specific distributions, the method improves diagnostic accuracy across heterogeneous datasets.
RankGraph-2 is a lifecycle co-design framework for billion-node graph learning in recommendation. It jointly optimizes graph construction, representation learning, and real-time serving, and introduces a learned cluster index to enable scalable retrieval without online KNN.
ScholarSum presents a student-teacher abstractive summarization framework guided by knowledge graph reasoning and reflective refinement. The system combines structured reasoning with iterative feedback to improve factual faithfulness and coherence in scientific paper summaries.