← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 22 papers for 2026-03-26

LineMVGNN: Anti-Money Laundering with Line-Graph-Assisted Multi-View Graph Neural Networks
GNN Graph Learning

LineMVGNN proposes a line-graph-assisted multi-view GNN for anti-money laundering on transaction digraphs. It addresses core limitations of conventional spectral GNNs by supporting multi-dimensional edge features, improving interpretability, and enhancing scalability. The approach shows stronger detection of suspicious transactions in AML tasks.

The impact of sensor placement on graph-neural-network-based leakage detection
GNN Graph Learning

This work investigates how sensor placement affects GNN-based leakage detection in water distribution networks. It introduces a PageRank-Centrality-based sensor placement strategy and demonstrates that sensor configuration substantially impacts detection accuracy. The results highlight the importance of strategic sensing for reliable leakage detection.

KCLNet: Electrically Equivalence-Oriented Graph Representation Learning for Analog Circuits
Graph Learning

KCLNet presents an electrically equivalent-oriented graph representation learning framework for analog circuits, grounded in DC equivalence and Kirchhoff’s current law constraints. It learns representations that reflect continuous analog behavior to support tasks like testability analysis and logical reasoning. The method utilizes asynchronous graph processing to capture circuit dynamics.

Reservoir-Based Graph Convolutional Networks
GNN Graph Learning

Reservoir-Based Graph Convolutional Networks introduce a reservoir computing-inspired approach to graph data, enabling deeper architectures without over-smoothing. A fixed random graph reservoir aggregates features across multiple hops, improving detection of long-range dependencies while maintaining efficiency. Experiments show gains on challenging graph tasks.

Cost-Sensitive Neighborhood Aggregation for Heterophilous Graphs: When Does Per-Edge Routing Help?
GNN Graph Learning

Cost-Sensitive Neighborhood Aggregation (CSNA) introduces a GNN layer that routes messages based on learned pairwise distances, separating concordant and discordant edges. This clarifies when per-edge routing helps in heterophilous graphs versus using a uniform spectral channel. The framework demonstrates improved performance in regimes where edge-level routing is informative.

Knowledge-Guided Manipulation Using Multi-Task Reinforcement Learning
Knowledge Graph

Knowledge-Guided Manipulation (KG-M3PO) couples knowledge graphs with multi-task model-based reinforcement learning for robotic manipulation in partially observable environments. It augments perception with an online 3D scene graph that grounds open-vocabulary detections into a relational representation, and trains a graph neural encoder end-to-end within RL.

CGRL: Causal-Guided Representation Learning for Graph Out-of-Distribution Generalization
Graph Learning

CGRL introduces causal-guided representation learning to improve GNN generalization under distribution shifts. A causal graph is constructed for node classification, and backdoor adjustment blocks non-causal paths to achieve robust out-of-distribution performance. The approach aims to reduce reliance on spurious correlations.

Dynamic Fusion-Aware Graph Convolutional Neural Network for Multimodal Emotion Recognition in Conversations
GNN Graph Learning

Dynamic Fusion-Aware GCN is proposed for multimodal emotion recognition in conversations, modeling interactions between speakers and learning dynamic fusion across modalities. Unlike fixed fusion schemes, it adapts to different emotion types, improving accuracy on MERC tasks.

Graph-Aware Late Chunking for Retrieval-Augmented Generation in Biomedical Literature
Graph Learning

Graph-Aware Late Chunking (GraLC-RAG) enhances retrieval-augmented generation for biomedical literature by integrating late chunking with graph-aware structural understanding. It surfaces evidence from across a document’s structural sections, not just top-ranked chunks, improving grounding and coverage.

CLiGNet: Clinical Label-Interaction Graph Network for Medical Specialty Classification from Clinical Transcriptions
GNN Graph Learning

CLiGNet is a clinical label-interaction graph network for medical specialty classification from clinical transcriptions, plus a leakage-free benchmark that reveals SMOTE leakage when splitting. It leverages interactions among labels to boost accuracy across 40 specialties on a robust, leakage-free dataset.

Conformal Risk Control for Safety-Critical Wildfire Evacuation Mapping: A Comparative Study of Tabular, Spatial, and Graph-Based Models
Graph Learning

Conformal Risk Control for wildfire evacuation mapping applies conformal risk control to wildfire spread prediction to provide formal guarantees on missed fire spread, and compares tabular, spatial, and graph-based models to highlight safety implications for evacuation planning.

Q-AGNN: Quantum-Enhanced Attentive Graph Neural Network for Intrusion Detection
GNN Graph Learning

Q-AGNN proposes a Quantum-Enhanced Attentive Graph Neural Network for intrusion detection, modeling network flows as nodes and edges representing similarity relations, and using quantum-inspired attention to improve detection performance.

Symbolic Graph Networks for Robust PDE Discovery from Noisy Sparse Data
GNN Graph Learning

Symbolic Graph Networks for Robust PDE Discovery from Noisy Sparse Data presents a symbolic graph framework for discovering partial differential equations under noisy and sparse conditions, avoiding reliance on local differentiation and improving robustness.

GraphRAG for Engineering Diagrams: ChatP&ID Enables LLM Interaction with P&IDs
Knowledge Graph Graph Learning

GraphRAG for Engineering Diagrams: ChatP&ID enables grounded and cost-effective LLM interaction with P&IDs by integrating graph retrieval augmented generation for engineering diagrams.

Avoiding Over-smoothing in Social Media Rumor Detection with Pre-trained Propagation Tree Transformer
GNN Graph Learning

Avoiding Over-smoothing in Social Media Rumor Detection with Pre-trained Propagation Tree Transformer analyzes how rumor propagation trees cause over-smoothing, and proposes a propagation-tree transformer to better capture long-range dependencies.

Can Graph Foundation Models Generalize Over Architecture?
GNN Graph Learning

Can Graph Foundation Models Generalize Over Architecture? discusses the architectural limitations of current graph foundation models, arguing that fixed backbones hinder true architecture-agnostic generalization and calling for more flexible designs.

Towards Intelligent Geospatial Data Discovery: a knowledge graph-driven multi-agent framework powered by large language models
Knowledge Graph Graph Learning

Towards Intelligent Geospatial Data Discovery presents a knowledge graph-driven multi-agent framework powered by large language models, addressing distributed, heterogeneous, and semantically inconsistent geospatial data landscapes to enable intelligent discovery.

MS-DGCNN++: Multi-Scale Dynamic Graph Convolution with Scale-Dependent Normalization for Robust LiDAR Tree Species Classification
GNN Graph Learning

MS-DGCNN++ introduces scale-dependent normalization for multi-scale dynamic graph convolution to handle varying LiDAR point densities, improving robustness in LiDAR-based tree species classification.

VL-KnG: Persistent Spatiotemporal Knowledge Graphs from Egocentric Video for Embodied Scene Understanding
Knowledge Graph Graph Learning

VL-KnG builds persistent spatiotemporal knowledge graphs from egocentric video for embodied scene understanding, processing video in chunks to maintain object identity via LLM-based associations, linking fine-grained scene graphs with global topologies without 3D reconstruction.

MemReward: Graph-Based Experience Memory for LLM Reward Prediction with Limited Labels
Graph Learning

MemReward introduces graph-based experience memory to predict LLM rewards under limited labels, leveraging remembered experiences to improve RL reward estimation and reduce reliance on expensive human labeling.

Mixture of Demonstrations for Textual Graph Understanding and Question Answering
Graph Learning LLM × Graph

Mixture of Demonstrations for Textual Graph Understanding and Question Answering introduces MixDemo, a GraphRAG framework that uses a mix of demonstrations to improve reasoning and answer accuracy in domain-specific QA. To mitigate degradation caused by irrelevant information in retrieved subgraphs, MixDemo emphasizes selecting high-quality demonstrations and handling noisy subgraphs. The approach aims to enhance text-based graph retrieval-augmented generation beyond zero-shot settings.

S-Path-RAG: Semantic-Aware Shortest-Path Retrieval Augmented Generation for Multi-Hop Knowledge Graph Question Answering
Knowledge Graph Graph Learning LLM × Graph

S-Path-RAG proposes a semantic-aware shortest-path Retrieval-Augmented Generation framework for multi-hop knowledge graph question answering. It enumerates bounded-length, semantically weighted candidate paths using a hybrid of weighted k-shortest, beam search, and constrained random-walk strategies. The model learns a differentiable path scorer, a contrastive path encoder, and a lightweight verifier, and injects a compact soft mixture of the selected path latents into the generation process.