Showing 29 papers for 2026-04-01
The authors establish a tight expressivity hierarchy for GNNs used in entity resolution in master data management, showing that for a given matching criterion a minimal MPNN architecture can be provably sufficient and that many extensions may be unnecessary overhead. They analyze which GNN components (message passing variants, port numbering, ego IDs) are essential, guiding the design of efficient ER models.
GSR-GNN introduces a training framework for deep GNNs on circuit graphs that reduces memory and compute with grouped-sparse-reversible techniques. It enables training extremely deep GNNs (hundreds of layers) and shows empirical gains over shallow models in circuit analysis tasks.
To mitigate oversquashing, they propose cross-attentive cohesive subgraph embedding that enriches node representations via cohesive subgraphs and cross-attention, alleviating information bottlenecks in long-range dependencies. This improves performance especially in dense/heterophilic regions.
CrossHGL is a text-free foundation model for cross-domain heterogeneous graph learning. It targets cross-domain generalization without relying on shared schemas or rich text attributes, offering a text-free foundation for heterogeneous graphs.
TMTE tackles multimodal attributed graphs by jointly evolving task-aware modality weights and topology. It addresses quality issues such as noisy interactions and missing connections, proposing a co-evolution framework that adapts both modalities and graph structure to downstream tasks.
FedDES proposes graph-based dynamic ensemble selection for personalized federated learning. Clients dynamically select peer contributions rather than average all peers, enabling instance-level personalization and reducing negative transfer.
Graph Vector Field (GVF) models health risk as a vector-valued field on time-varying simplicial complexes, coupling discrete differential geometry with modality-structured mixture-of-experts. Risk is represented as a vector-valued cochain evolved by Hodge Laplacians.
ORACAL is a robust multimodal framework for smart contract vulnerability detection with causal graph enrichment. It augments GNN-based detection with causal-graph insights to explain evidence and improve robustness against adversarial manipulation.
NeiGAD augmenting GAD via Spectral Neighbor Information. It is a plug-and-play method that explicitly models spectral neighbor information to improve anomaly detection in graphs, achieving higher accuracy and robustness.
FairGC introduces fairness-aware graph condensation, addressing bias in synthetic graphs used to scale GNNs. It preserves utility while enforcing fairness, avoiding amplification of demographic disparities.
Contextual Graph Representations for Task-Driven 3D Perception and Planning. It proposes task-focused contextual graphs that prune large 3D scene graphs to subgraphs relevant to a given task, enabling efficient perception and planning.
Graph Attention Network-Based Detection of Autism Spectrum Disorder. The study builds a GAT-based classifier (GATGraphClassifier) on fMRI-derived functional connectivity networks from ABIDE, achieving improved ASD detection results.
Continual Graph Learning: A Survey. This survey reviews methods for continual learning on graphs, including experience replay and generative replay, and discusses challenges such as information preservation, privacy, and distribution shifts.
Towards Quantifying Long-Range Interactions in Graph Machine Learning: a Large Graph Dataset and a Measurement. The authors introduce City-Networks, a large-scale dataset and measurement protocol to quantify long-range dependencies in graph ML, enabling systematic evaluation of models’ capacity to capture distant dependencies.
Compact Conformal Subgraphs. This work introduces graph-based conformal compression to produce compact subgraphs that preserve statistical validity. It formulates compression as selecting a smallest subgraph capturing a prescribed probability mass and reduces it to a weighted densest k-subgraph problem.
Graph-Aware Stealthy Poison-Text Backdoors for Text-Attributed Graphs. This work studies a realistic attack where adversaries inject malicious cues only into node text while leaving the graph structure unchanged, enabling graph-aware backdoors in text-augmented graphs.
PEANUT: Perturbations by Eigenvector Alignment for Attacking Graph Neural Networks Under Topology-Driven Message Passing. The paper presents a gradient-free, restricted black-box attack that perturbs eigenvector alignments to degrade GNNs, using virtual nodes to influence topology-driven message passing.
LSM-GNN: Large-scale Storage-based Multi-GPU GNN Training by Optimizing Data Transfer Scheme. The framework uses storage-based out-of-core data handling and optimized inter-GPU transfers to scale GNN training to very large graphs while reducing memory bottlenecks.
A Survey of Heterogeneous Graph Neural Networks for Cybersecurity Anomaly Detection. This survey covers HGNN-based approaches for cybersecurity anomaly detection, highlighting heterogeneous graphs, temporal dynamics, and practical challenges.
AutoRegressive Generation with B-rep Holistic Token Sequence Representation. The paper proposes BrepARG, a holistic token sequence representation that encodes geometry and topology of B-rep into a single sequence, enabling sequence-based B-rep generation with transformers.
The paper proposes Efficient and Scalable Granular-ball Graph Coarsening Method for Large-scale Graph Node Classification. It tackles the high computational overhead of deep GCNs on large graphs by employing a granular-ball coarsening approach that preserves multi-granularity information to enable scalable training and accurate classification.
The paper presents VulGNN, a lightweight graph neural network for software vulnerability detection, leveraging the natural graph structure of code. It achieves performance close to large language models while being about 100× smaller and much faster to retrain.
This work evaluates the effectiveness of Large Language Models (LLMs) for automatically generating RDF knowledge graphs from cloud logs. It analyzes multiple architectures and discusses their performance and limitations on complex cloud log data.
CLAUSE introduces a three-agent neuro-symbolic framework that treats context construction as a sequential decision process over knowledge graphs, deciding what to expand and which paths to follow to balance accuracy, latency, and provenance. It aims to reduce over-retrieval and unpredictable runtimes in multi-hop KG reasoning.
UniAI-GraphRAG enhances GraphRAG with ontology-guided extraction, multi-dimensional clustering, and dual-channel fusion to improve robustness and cross-domain adaptability in multi-hop reasoning.
CSNA introduces a cost-sensitive neighborhood aggregation layer for heterophilous graphs, using learned projections to soft-route messages through concordant and discordant edges; it analyzes when per-edge routing helps versus when a uniform spectral channel suffices.
EvidenceNet provides a framework and dataset to build disease-specific knowledge graphs from full-text biomedical literature, using an LLM-assisted pipeline to extract experimentally grounded findings as structured evidence nodes, normalize entities, score evidence quality, and connect evidence records with provenance.
UltRAG presents a universal, simple, scalable recipe for knowledge graph RAG, addressing hallucination by grounding retrieval in the graph and enabling reliable multi-hop reasoning over KG.
PhaseGraph calibrates heterogeneous graph-vector retrieval signals for multi-hop QA by mapping vector similarities and graph-based scores to a common percentile-based scale, enabling stable fusion without discarding magnitude information.