Showing 13 papers for 2026-05-01
Context-Aware Graph Attention for Unsupervised Telco Anomaly Detection proposes C-MTAD-GAT, an unsupervised context-aware graph-attention model for anomaly detection in multivariate time series from mobile networks. It blends graph attention with lightweight context embeddings and uses a deterministic reconstruction head plus a multi-step forecaster to produce anomaly scores. Thresholds are calibrated without labels from validation residuals, keeping the whole pipeline unsupervised. On TELCO data, it outperforms MTAD-GAT and Telco-specific baselines.
TypeBandit: Type-Level Context Allocation and Reweighting for Effective Attribute Completion in Heterogeneous Graph Neural Networks proposes a lightweight, model-agnostic approach to attribute completion in heterogeneous graphs by addressing type-dependent information asymmetry. It introduces type-level context allocation and reweighting guided by topology-aware initialization to better exploit signals from each node type. The method is designed to plug into existing GNN pipelines and improve missing attribute imputation across heterogeneous graphs.
Toward Scalable SDN for LEO Mega-Constellations: A Graph Learning Approach proposes a scalable, hierarchical software-defined networking framework for LEO mega-constellations. The approach uses graph neural networks to compactly represent the constellation topology and inter-satellite links, enabling efficient control and routing decisions. The results demonstrate improved scalability and manageability for future large satellite networks.
On the Expressive Power of GNNs to Solve Linear SDPs investigates how expressive power in graph neural networks can capture the optimal solutions of linear semidefinite programs (SDPs). The work characterizes conditions under which GNNs can recover SDP optima and discusses the limits of graph-based surrogates for convex optimization. The results help understand when a GNN is a suitable surrogate for SDP solvers.
Graph-Based Biomarker Discovery and Interpretation for Alzheimer's Disease proposes a graph-based framework to identify blood-based biomarkers for Alzheimer's disease and interpret their role as potential drug targets. It leverages accessible blood tests to enable population-scale screening, addressing limitations of radiological imaging. The approach integrates molecular and clinical signals in a graph to reveal candidate biomarkers and pathways.
Efficient Traffic Forecasting on Large-Scale Road Network by Regularized Adaptive Graph Convolution addresses scalability of spatial-temporal graph convolution for large road networks by introducing a regularized adaptive graph convolution. The approach reduces quadratic complexity while preserving predictive performance, enabling accurate traffic forecasting on large graphs. Experiments show improved scalability with competitive accuracy.
AEGIS: Authentic Edge Growth In Sparsity for Link Prediction in Edge-Sparse Bipartite Knowledge Graphs introduces AEGIS, an edge-only augmentation framework for link prediction in edge-sparse bipartite knowledge graphs. It resamples existing training edges with either uniform or inverse-degree bias to avoid fabricating new endpoints, preserving the original node set. The framework enables robust link prediction across sparse regimes and natural sparse graphs.
Evaluating Assurance Cases as Text-Attributed Graphs for Structure and Provenance Analysis proposes a graph diagnostic framework to analyze assurance cases by modeling them as text-attributed graphs. The framework focuses on link prediction to uncover connections between argument elements and graph classification to differentiate between assurance case structures. This enables systematic analysis of structure, traceability, and provenance in regulated settings.
OptimusKG: Unifying biomedical knowledge in a modern multimodal graph introduces OptimusKG, a multimodal labeled property graph for biomedical knowledge built from structured and semi-structured resources to preserve type-specific metadata across molecular, anatomical, clinical, and environmental domains. It contains over 190,000 nodes and aims to enable integrated reasoning across diverse biomedical domains. The graph supports downstream tasks such as retrieval and reasoning.
Robust Learning on Heterogeneous Graphs with Heterophily: A Graph Structure Learning Approach studies robust representation learning on heterogeneous graphs with heterophily, where connected nodes may have different labels. It identifies structural noise as a key challenge that degrades performance and proposes a graph structure learning method to automatically identify and mitigate noisy connectivity, improving robustness. Experiments demonstrate gains under noisy conditions.
Knowledge Graph Representations for LLM-Based Policy Compliance Reasoning constructs knowledge graphs from AI risk-related policy documents and uses them to answer questions with LLMs. The study evaluates five LLMs on 42 policy QA tasks spanning six reasoning types, demonstrating how structured KG representations support policy compliance reasoning. It uses two ontology schemas to organize policy knowledge.
LLM as Clinical Graph Structure Refiner: Enhancing Representation Learning in EEG Seizure Diagnosis proposes using large language models to refine the graph structure used for EEG-based seizure diagnosis. Because EEG graphs are noisy, LLM-guided edge selection helps prune irrelevant or redundant connections, improving representation learning for seizure detection. The approach aims to leverage LLM reasoning to improve graph quality for clinical EEG analysis.
From Unstructured to Structured: LLM-Guided Attribute Graphs for Entity Search and Ranking presents a two-stage approach that uses large language models to convert unstructured product descriptions into structured attribute graphs for entity search. In the offline stage, it extracts structured attributes from text, then constructs reusable attribute graphs, enabling context-aware similarity and ranking. A graph-aware LLM ranking component further improves search quality.