← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 16 papers for 2026-04-24

TRAVELFRAUDBENCH: A Configurable Evaluation Framework for GNN Fraud Ring Detection in Travel Networks
GNN Graph Learning

TravelFraudBench (TFG) is a configurable benchmark for evaluating GNNs on fraud ring detection in travel platform graphs. It overcomes limitations of existing benchmarks by simulating three travel-specific ring topologies: ticketing fraud with star patterns on device/IP clusters, ghost hotel schemes formed by reviewer-hotel bipartite cliques, and account takeover rings. This enables cross-topology evaluation of GNN detectors across structurally distinct fraud rings.

Spectral Embeddings Leak Graph Topology: Theory, Benchmark, and Adaptive Reconstruction
GNN Graph Learning Graph Theory

We introduce LoGraB, a Local Graph Benchmark, a unified framework to evaluate GNNs under fragmented, privacy-sensitive graphs. It decomposes standard datasets into fragmented benchmarks using three strategies and four controls, including neighborhood radius d. The framework also analyzes how spectral embeddings can leak topology and provides adaptive reconstruction benchmarks.

Graph Neural Network-Informed Predictive Flows for Faster Ford-Fulkerson and PAC-Learnability
GNN Graph Learning Graph Theory

We propose a learning-augmented framework that speeds up max-flow computations and image segmentation by combining GNNs with Ford-Fulkerson. Instead of predicting initial flows, the model learns edge-importance probabilities to guide augmenting path selection. The proposed MPGNN learns node and edge embeddings via message passing, capturing global structure and local dynamics like residual capacity and bottlenecks.

Drug Synergy Prediction via Residual Graph Isomorphism Networks and Attention Mechanisms
GNN Graph Learning

The paper tackles drug synergy prediction using residual graph isomorphism networks and attention mechanisms. By modeling drugs and their interactions as graphs, the method uses residual GINs and attention to capture structural and relational information, improving synergy score prediction. This aims to reduce experimental validation costs.

StormNet: Improving storm surge predictions with a GNN-based spatio-temporal offset forecasting model
GNN Graph Learning

StormNet is a spatio-temporal GNN for bias correction of storm surge forecasts. It integrates a GNN-based offset correction to existing forecasts (e.g., ADCIRC) to reduce biases and uncertainties due to model and data limitations. It leverages spatial-temporal dependencies to forecast corrections.

GARG-AML against Smurfing: A Scalable and Interpretable Graph-Based Framework for Anti-Money Laundering
Graph Learning

GARG-AML is a fast, transparent graph-based framework to detect smurfing in money laundering. It assigns a single, easy-to-interpret risk score to every account in both directed and undirected networks. It constructs an adjacency matrix from second-order neighborhoods and measures block densities to identify suspicious patterns.

Forecasting Individual NetFlows using a Predictive Masked Graph Autoencoder
Graph Learning

The paper presents a proof-of-concept GNN model to predict NetFlow at the flow level. It uses sliding windows to partition heterogeneous bidirectional graphs with IP, port, and connection nodes, and uses GNN to model evolution of graph structure and features. The method shows superior performance in identifying the IPs and ports involved in traffic.

SkillGraph: Graph Foundation Priors for LLM Agent Tool Sequence Recommendation
LLM × Graph

SkillGraph is a tool to improve LLM agent tool sequencing by learning from trajectories. It builds a directed weighted execution-transition graph that encodes workflow precedence, providing reusable graph priors for tool selection and ordering beyond semantic similarity. It mitigates ordering errors in structured domains.

Large Language Models Meet Biomedical Knowledge Graphs for Mechanistically Grounded Therapeutic Prioritization
LLM × Graph Knowledge Graph

DrugKLM combines biomedical knowledge graphs with LLM-based mechanistic reasoning to prioritize therapeutics mechanistically. It uses KG structure plus LLM to rank candidates; outperforms KG-only and LLM-only baselines, including TxGNN.

AutoGraph-R1: End-to-End Reinforcement Learning for Knowledge Graph Construction
Knowledge Graph Graph Learning

AutoGraph-R1 trains an LLM constructor with reinforcement learning to directly optimize KG construction for retrieval-augmented generation QA. It bridges KG construction and downstream task performance by RL, enabling end-to-end optimization.

On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence
Graph Learning

On-Meter Graph ML is a case study of using graph ML for PV power forecasting at the grid edge with smart meters. It describes hardware/software, ONNX deployment, and ML models GCN and GraphSAGE, plus a custom ONNX operator. Emphasizes edge intelligence.

ItemRAG: Item-Based Retrieval-Augmented Generation for LLM-Based Recommendation
Knowledge Graph Graph Learning

ItemRAG is item-based retrieval-augmented generation for LLM-based recommendation. It uses user purchase histories and similar items to retrieve relevant information to feed LLM for better recommendations; addresses noise and cold-start.

Explainable Iterative Data Visualisation Refinement via an LLM Agent
Graph Learning

Explainable Iterative Data Visualization Refinement via an LLM Agent proposes an LLM agent to iteratively refine visualizations, selecting algorithms and hyperparameters to produce faithful plots, with explanations and justification.

EgoSelf: From Memory to Personalized Egocentric Assistant
Graph Learning

EgoSelf: From Memory to Personalized Egocentric Assistant introduces a graph-based interaction memory from past observations and a dedicated learning task for personalization to tailor egocentric assistants to individual users.

Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis
Graph Learning

Caesar: Deep Agentic Web Exploration for Creative Answer Synthesis presents an agentic LLM architecture that uses a knowledge graph to foster associative reasoning, enabling creative synthesis rather than passive retrieval by exploiting structured web knowledge.

Deep Interest Mining with Cross-Modal Alignment for SemanticID Generation in Generative Recommendation
Graph Learning

Deep Interest Mining with Cross-Modal Alignment for SemanticID Generation in Generative Recommendation addresses information degradation and semantic degradation in two-stage compression, proposing cross-modal alignment and deep interest mining to produce higher-quality Semantic IDs for generative recommendations.