← Home

Daily arXiv Papers

Graph Neural Networks · Graph Learning · LLM × Graph

Showing 13 papers for 2026-03-27

DyMRL: Dynamic Multispace Representation Learning for Multimodal Event Forecasting in Knowledge Graph
Knowledge Graph

DyMRL proposes Dynamic Multispace Representation Learning for multimodal knowledge graphs to improve event forecasting. It tackles time-sensitive and dynamic structural information across modalities, aiming to fuse multimodal knowledge as it evolves. The approach seeks to capture deeper interactions beyond static or shallow dynamic models.

Learning Mesh-Free Discrete Differential Operators with Self-Supervised Graph Neural Networks
GNN Graph Learning

We introduce a parametrised framework to learn mesh-free discrete differential operators with a graph neural network trained under polynomial moment constraints from truncated Taylor expansions. The model maps local stencil relative positions directly to operator weights, enabling accurate operators with low per-stencil cost. This approach improves accuracy without relying on mesh-based discretizations.

GraphER: An Efficient Graph-Based Enrichment and Reranking Method for Retrieval-Augmented Generation
Graph Learning

GraphER introduces a graph-based enrichment and reranking strategy to improve retrieval in RAG systems when evidence is dispersed across sources. It exploits the data's organizational structure to connect related pieces of evidence and rank them more effectively, reducing reliance on iterative query expansions. The method aims to yield more relevant supporting evidence for generated answers.

Dual-Graph Multi-Agent Reinforcement Learning for Handover Optimization
GNN Graph Learning

We present a dual-graph multi-agent reinforcement learning framework to optimize handover control in cellular networks. By modeling pairwise interactions of Cell Individual Offsets (CIO) across neighboring cells on two graphs, the approach captures coupled effects and enables coordinated CIO tuning. The method improves mobility management and network performance under non-stationary traffic conditions.

Autotuning T-PaiNN: Enabling Data-Efficient GNN Interatomic Potential Development via Classical-to-Quantum Transfer Learning
GNN Graph Learning

We introduce Transfer-PaiNN (T-PaiNN), a data-efficient framework for training GNN-based interatomic potentials. By transferring knowledge from inexpensive classical force fields to quantum-level targets, it substantially reduces the quantum data required while retaining accuracy close to DFT benchmarks.

Morphling: Fast, Fused, and Flexible GNN Training at Scale
GNN Graph Learning

Morphling tackles the hardware challenge of training GNNs by fusing irregular graph traversals with regular dense matrix operations. It provides a fast, scalable training pipeline with optimized kernels, improved cache locality, reduced memory movement, and fewer intermediate allocations compared to existing frameworks.

OWLEYE: Zero-Shot Learner for Cross-Domain Graph Data Anomaly Detection
GNN Graph Learning

OWLEYE proposes a zero-shot learner for graph anomaly detection that generalizes across domains to detect anomalies in unseen graphs without retraining. It leverages cross-domain knowledge and self-supervision to identify anomalous patterns in diverse graph datasets.

Position: Spectral GNNs Are Neither Spectral Nor Superior for Node Classification
GNN Graph Learning

We argue that Spectral GNNs are not true spectral methods nor consistently superior for node classification. Two main issues are identified: graph Fourier bases are not classical Fourier bases for graph signals, and (n-1)-degree polynomials can exactly interpolate any spectral filter, undermining typical spectral explanations. The paper suggests rethinking spectral approaches and exploring alternatives.

Interactive Query Answering on Knowledge Graphs with Soft Entity Constraints
Knowledge Graph

We propose interactive query answering on incomplete knowledge graphs with soft entity constraints, enabling queries that incorporate vague or context-dependent preferences. By allowing soft constraints and user interaction, the framework yields more flexible and user-aligned answers beyond strict first-order logic queries.

UniAI-GraphRAG: Synergizing Ontology-Guided Extraction, Multi-Dimensional Clustering, and Dual-Channel Fusion for Robust Multi-Hop Reasoning
Graph Learning Knowledge Graph

UniAI-GraphRAG enhances retrieval-augmented reasoning by combining ontology-guided knowledge extraction, multi-dimensional clustering, and dual-channel fusion on GraphRAG. These components improve cross-domain adaptability, data integrity, and multi-hop reasoning performance.

Evaluating adaptive and generative AI-based feedback and recommendations in a knowledge-graph-integrated programming learning system
Knowledge Graph

We design and evaluate a framework that integrates a large language model with retrieval-augmented generation and a knowledge graph into an adaptive programming learning system. It assesses code, provides formative feedback, and recommends exercises, comparing adaptive, GenAI, and hybrid GenAI-adaptive modes from learner perspectives.

Graph-of-Mark: Promote Spatial Reasoning in Multimodal Language Models with Graph-Based Visual Prompting
Graph Learning

Graph-of-Mark extends visual prompting by representing marked image regions as a graph to capture relations among objects, enabling better spatial grounding for multimodal language models. It moves beyond isolated region marks to reason about spatial structure.

A Hypergraph-Based Framework for Exploratory Business Intelligence
Graph Theory

ExBI introduces a hypergraph-based BI framework with Source, Join, and View operators to enable dynamic schema evolution and materialized view reuse. It supports exploratory, iterative analytics and reduces reliance on expert knowledge and costly recomputation.