Showing 35 papers for 2026-05-05
PhaseNet++ proposes phase-aware anomaly detection in the frequency domain for industrial control systems. It argues that phase information from time-frequency transforms provides a complementary detection modality to amplitude-based features and leverages phase coherence graphs to model inter-sensor relationships for anomaly detection.
This survey reviews graph rewiring techniques to mitigate over-squashing and over-smoothing in GNNs. It discusses topology-aware rewiring methods, provides a taxonomy of approaches, and highlights open challenges and future research directions.
TopoNTK introduces an infinite-width kernel for simplicial message passing on edge features, enabling higher-order interactions beyond pairwise graphs. It combines lower and upper Hodge interactions to capture multi-way relationships in relational systems.
We study whether spectral graph sparsification preserves representation geometry in polynomial-filter GNNs. The main result shows that an epsilon-spectral sparsifier induces only O(epsilon) perturbations in polynomial graph filters and in the learned representations across layers, preserving the overall geometry.
This paper rethinks multi-label node classification by asking whether carefully tuned classic full-graph GNNs can match specialized label-aware models. Through systematic experiments, it shows strong baseline performance for well-tuned GNNs, challenging the need for complex label-aware designs.
Evaluating LLMs on large-scale graph property estimation via random walks to probe reasoning abilities on graph algorithms. The study benchmarks how well LLMs can infer graph properties and perform related graph-analytic reasoning tasks.
Identifies bottlenecks in ML surrogates for CFD, noting that node-wise supervision and explicit Euler time-stepping ignore stiff dynamics and local flux continuity. Proposes incorporating spatial-temporal awareness in mesh-based surrogates to address these gaps.
FedTGNN-SS is a privacy-preserving federated semi-supervised GNN framework with prototype-guided pseudo-labeling for gestational diabetes prediction. It enables cross-hospital learning with limited labels while protecting patient privacy.
This work studies differential privacy in GCNs via subsampling stability, deriving upper bounds on misclassification rate as a function of the subsampling probability ps. It analyzes the privacy-utility trade-off and identifies feasible ranges of ps for meaningful guarantees.
This paper introduces a large-margin classifier with graph-based adaptive regularization on Gabriel graphs. By using per-class regularization hyperparameters, it better handles outliers and class imbalance by shaping margins and decision thresholds.
TIJERE presents a novel threat intelligence joint extraction model that leverages analyst expert knowledge to jointly extract entities and relations from threat reports, addressing feature confusion, language ambiguity, and noise propagation.
H3 proposes a healthcare three-hop index to predict physician referral networks, capturing indirect pathways through intermediate nodes to overcome sparsity, disassortative degree mixing, and hub-dominated topology.
This spatial-temporal learning-based distributed routing framework for dynamic LEO satellite networks combines GAT, LSTM, and DQN within a POMDP framework to enable distributed routing decisions based on local observations.
Middle-mile logistics is reformulated as a multi-objective goal-conditioned MDP, integrating graph neural networks with model-free reinforcement learning to extract compact feature graphs from the environment for routing parcels.
PIEGraph couples analytical physics with graph neural networks to learn equivariant object dynamics from few interactions, enabling accurate modeling of both rigid and deformable bodies with improved data efficiency and long-horizon feasibility.
Fine-Grained Graph Generation through Latent Mixture Scheduling introduces a conditional variational autoencoder that dynamically aligns graph and property representations to enable fine-grained control over generated graphs.
GraphLand evaluates graph machine learning models on diverse industrial data, arguing that current benchmarks cover limited domains and advocating evaluation of graph foundation models across varied datasets for robust transferability.
Effective Capacitance Modeling Using Graph Neural Networks applies GNNs to predict effective capacitance for VLSI timing analysis, enabling improved early-stage timing predictions and better placement/routing decisions.
Text-Attributed Knowledge Graph Enrichment with Large Language Models for Medical Concept Representation explores using LLMs to enrich medical concept representations with textual attributes, addressing missing cross-type dependencies in ontology resources.
On the Expressive Power of GNNs to Solve Linear SDPs investigates what graph-based capacity is sufficient for recovering optimal SDP solutions, providing theoretical insights into the capabilities of GNN-based surrogates.
Ligandformer introduces a multi-layer self-attention graph neural network for predicting compound properties with interpretable attention. The model provides robust interpretation of QSAR predictions, enabling validation of explanations by chemists and offering heuristic guidance for structure optimization.
This paper studies backdoor attacks on graphs that transfer across learning paradigms. It introduces promptable subgraph triggers to enhance cross-paradigm transferability and attack success across graph supervised, contrastive, and prompt-based learning, and evaluates the effectiveness under diverse GNN frameworks.
We define recurrent GNNs as Recurrent Arithmetic Circuits (RACs), generalizing beyond standard GNN architectures. RACs behave like memory-enabled arithmetic circuits, with memory gates storing information across iterations, enabling richer computation. We analyze their computational power and expressiveness relative to other models.
HyperODE RCA combines hypergraph attention, latent ODEs, and multimodal cross-attention to localize root causes in microservice systems. It learns higher-order service interactions via differentiable hyperedges and models irregular observation dynamics with latent ODEs, enabling fine-grained RCA.
Introduces C-MTAD-GAT, a context-aware multivariate time-series graph attention model for unsupervised anomaly detection in large-scale mobile networks. It handles thousands of KPI time series, supports a single shared model, and is robust to context shifts and nonstationarity without labeled data.
Proposes Decoupled Relation Alignment to empower heterogeneous Graph Foundation Models on multi-domain heterogeneous graphs. Unlike global feature alignment, which distorts type-specific semantics, the approach decouples relation alignment from domain features to avoid Type Collapse and Relation Confusion, preserving topology while aligning cross-domain information.
We introduce controllable hypothesis generation for abductive reasoning on knowledge graphs to reduce redundant or irrelevant hypotheses. The task adds controllability constraints and ranking incentives to steer generation toward relevant, high-quality hypotheses, enabling practical use in domains like clinical diagnosis and scientific discovery.
G-reasoner presents a foundation-model-based system for unified reasoning over graph-structured knowledge. It addresses limitations of static parametric knowledge in LLMs by using retrieval-augmented reasoning and graph-aware structure modeling, integrating GraphRAG-style retrieval with graph reasoning for knowledge-intensive tasks.
This survey reviews graph rewiring techniques to mitigate over-squashing and over-smoothing in GNNs. It categorizes methods (e.g., skip connections, edge rewiring, topology-aware propagation), discusses theoretical insights, empirical outcomes, and challenges for scalable, robust graph learning.
Introduces Semantic Level of Detail (SLoD) for knowledge graphs, enabling continuous control of abstraction boundaries via spectral heat diffusion. It provides a principled framework for navigating graphs at different granularities with formal guarantees about abstraction transitions.
KG-First, LLM-Fallback introduces SkillGraph-Service, a hybrid microservice unifying competency resources into a provenance-preserving knowledge graph. It uses a KG-first approach with LLM fallback to support grounded skill search and explanation.
MMSC is a self-supervised, multi-modal relational item representation learning framework for inferring substitutable and complementary items. It leverages multi-modal content and behavior signals to overcome noisy supervision and long-tail data, producing robust item recommendations.
TagRAG introduces a tag-guided hierarchical retrieval-augmented generation framework for knowledge graphs. Tags guide retrieval and enable efficient, scalable graph-aware generation, addressing inefficiencies and incremental updates in GraphRAG.
UniQGen is a constraint-guided framework that uses LLM agents to extract and refine graph query clauses into executable queries across languages. It employs a variant of Chase & Backchase to ensure queries are correct and aligned with user intent.
U-HNSW proposes an efficient graph-based ANNS method for universal Lp metrics. A single index supports queries across all p in (0,2], achieving better query efficiency than MLSH, with performance advantages on diverse p-values.