Showing 39 papers for 2026-03-24
This paper introduces TAGBD, a text-only backdoor attack against text-attributed graphs. The attacker modifies only node text, leaving the graph structure intact, to trigger targeted mispredictions during deployment. The work also discusses detection and defense considerations for such backdoors.
We propose SDE-HGNN, an SDE-driven spatio-temporal hypergraph neural network for irregular longitudinal fMRI connectome modeling in Alzheimer's disease. An SDE-based Reconstruction module recovers continuous latent trajectories from irregular observations, and a hypergraph-based model captures complex spatio-temporal relations.
This work presents Spatio-Temporal Grid Intelligence, a hybrid framework that combines Graph Neural Networks and LSTMs for robust electricity theft detection. It leverages time-series anomaly detection, supervised learning, and graph modeling to better capture spatio-temporal patterns of non-technical losses. The result is improved detection performance under real-world dynamics.
We study adversarial attacks on locally private GNNs under Local Differential Privacy. The work analyzes how DP guarantees interact with adversarial perturbations, showing conditions under which privacy can either hinder or enable attacks, and discusses potential defenses. The study provides empirical and theoretical insights into robustness under LDP.
This survey offers a unified, industrially grounded framework for Attributed Graph Clustering (AGC), highlighting the gap between academic benchmarks and real-world deployment. It discusses challenges across modeling, evaluation, and deployment in industry and outlines practical directions for bridging the gap.
The paper argues that Riemannian geometry should guide next-generation Graph Foundation Models (GFMs). It discusses limitations of conventional GNNs in memory and interpretability and advocates geometric, non-Euclidean approaches to achieving more powerful, principled graph intelligence.
MISApp proposes multi-hop intent-aware session graph learning for next app prediction. By modeling higher-order dependencies and evolving session intent, it improves prediction beyond local sequential or single-hop transitions, particularly in cold-start scenarios.
FastPFRec introduces Fast Personalized Federated Recommendation with Secure Sharing. It accelerates training via an efficient local update strategy and strengthens data privacy with secure sharing, achieving faster convergence while mitigating privacy risks.
Multi-RF Fusion with Multi-GNN Blending achieves state-of-the-art molecular property prediction by ensembling 12 Random Forest models trained on concatenated fingerprints with deep GNN predictions, blended at 12% weight. This architecture demonstrates the value of combining traditional fingerprints with modern graph learning.
This work studies privacy-preserving graph structure learning for open graph data, focusing on privacy guarantees at the data publishing stage rather than training. It uses differential privacy-inspired mechanisms to balance privacy with utility in open datasets and discusses practical deployment considerations.
From Nodes to Narratives introduces GSPELL, a lightweight post-hoc explanation method that uses large language models and graph context to generate interpretable, narrative rationales for GNN predictions on text-attributed graphs. It aims to produce fine-grained explanations that are easy to understand.
The paper scales kinetic Monte-Carlo simulations of grain growth by a hybrid CNN-GNN architecture: a CNN-based bijective autoencoder compresses spatial dimensions and a GNN evolves the microstructure in latent space, enabling larger-scale simulations.
This work improves fairness of LLM-based ICU mortality prediction using case-based prompting. It analyzes demographic biases and demonstrates that case-based prompts can enhance fairness without substantially sacrificing accuracy.
The paper investigates geometric coherence in federated GNNs, showing a mismatch between standard Euclidean aggregation and operator-valued GNNs. Local updates perturb operator manifolds, challenging the effectiveness of global aggregation under heterogeneous client graphs.
Lost in Aggregation formalizes a fundamental expressivity limit of MP-GNNs: any MP-GNN with common aggregations induces only polynomially many equivalence classes, while the number of non-isomorphic graphs is doubly-exponential. This highlights inherent limitations in distinguishing graph structures with typical message-passing schemes.
LLM-Enhanced Energy Contrastive Learning advances out-of-distribution detection for text-attributed graphs by leveraging LLM-derived textual representations within an energy/contrastive learning framework to detect anomalous nodes.
Towards Intelligent Geospatial Data Discovery proposes a knowledge graph-driven multi-agent framework powered by LLMs to improve geospatial data discovery. The framework combines KG-based semantic reasoning and multi-agent collaboration to better interpret user intent and retrieve relevant data.
gUFO offers a gentle, lightweight implementation of the Unified Foundational Ontology for Semantic Web Knowledge Graphs, designed for OWL 2 DL applications and currently targeted for ISO standardization. It aims to provide a practical, interoperable foundational ontology.
KLDrive introduces a knowledge-graph-augmented LLM reasoning framework for fine-grained 3D scene reasoning in autonomous driving. It enables precise, graph-enabled reasoning over driving scene facts to improve QA tasks and reduce hallucinations.
The paper investigates universal adversarial attacks on text-attributed graphs, assessing whether LLMs and backbones can be fooled by a single, architecture-agnostic perturbation. It discusses the design, effectiveness, and potential defenses against universal attacks.
DomAgent proposes to integrate knowledge graphs and case-based reasoning to improve domain-specific code generation, addressing gaps that generic LLMs have when working with real-world software projects. By leveraging structured domain knowledge and past coding cases, it aims to produce more accurate, context-aware solutions.
Agentic Personas for Adaptive Scientific Explanations with Knowledge Graphs argues that explanations are often static and fail to adapt to expert goals, reasoning strategies, or decision contexts. By combining knowledge graphs with agentic personas, explanations can be tailored to different reasoning strategies, enabling grounded and diverse explanations in scientific discovery.
GSEM organizes clinical experiences into a dual-layer memory graph that captures both the decision structure within each experience and the relationships across experiences, enabling reuse and more reliable retrieval. This graph-based memory can reduce noisy retrieval and improve clinical reasoning compared with flat memories.
The work presents a multi-layer AI-driven framework to map how interdisciplinary teams converge on shared viewpoints over time. It uses LLMs to extract viewpoints aligned with a Needs-Approach-Benefits framework, and combines graph visualization with human-in-the-loop evaluation.
GEM is a native graph-based index designed for multi-vector representations, preserving multi-vector semantics. It overcomes limitations of reusing single-vector indexes and provides efficient retrieval.
The study presents graph-based data-driven discovery of interpretable laws governing corona-induced noise and radio interference for high-voltage transmission lines. By mining data with graph-based methods, it reveals interpretable laws to predict and manage AN and RI in ultra-high-voltage grids, supporting compliant deployment.
The paper introduces AEGIS, a graph-guided deep vulnerability reasoning system that uses dialectics and meta-auditing to ground evidence in a data-flow topology, mitigating ungrounded deliberation and spurious dependencies. By structuring clues into a graph and auditing it, it aims to yield more reliable vulnerability verdicts.
KG-Hopper enables compact open LLMs to perform knowledge graph reasoning through reinforcement learning, enabling multi-hop KBQA over structured graphs. It addresses the limitations of small models in complex KBQA tasks.
HECG introduces a hierarchical error-corrective graph framework for autonomous agents with LLM-based action generation. It uses Multi-Dimensional Transferable Strategy (MDTS), integrating Q, C, R, and LLM-Score for precise selection of high-quality strategies and reducing errors.
Knowledge Fusion via Bidirectional Information Aggregation proposes bidirectional information aggregation to fuse dynamic knowledge graphs with static LLMs, addressing knowledge staleness and forgetting without heavy parameter updates. The approach aims to keep LLMs current with KG knowledge.
LLM-Enhanced Rumor Detection via Virtual Node Induced Edge Prediction introduces an LLM-enabled framework where virtual nodes induce edge predictions to model propagation dynamics in rumor spread, integrating textual coherence with diffusion patterns for improved detection.
BubbleRAG addresses hallucinations in retrieval-augmented generation for black-box knowledge graphs by identifying three core challenges causing recall and precision issues, and proposes an evidence-driven retrieval augmented generation approach.
Ontology-Compliant Knowledge Graphs explores ontology-compliant KGs, offering schema-based explainability and interoperability, and introduces term-matching algorithms and pattern-based compliance metrics, with a building sector case study.
LSA: A Long-Short-term Aspect Interest Transformer for Aspect-Based Recommendation proposes a model that captures dynamic shifts in user interests for fine-grained recommendations, using graph-structured relations among users, items, and aspect terms.
Graph Fusion Across Languages using Large Language Models proposes a cross-lingual graph fusion framework that leverages in-context reasoning and multilingual priors, mapping KG triplets into natural language to reconcile entities and relations across languages.
StreamTGN: A GPU-Efficient Serving System for Streaming Temporal Graph Neural Networks introduces StreamTGN, a streaming inference system that exploits locality: a new edge affects only nodes within L hops of the endpoints, reducing updates and enabling GPU-friendly serving.
FGIM: a Fast Graph-based Indexes Merging Framework for Approximate Nearest Neighbor Search proposes merging multiple graph-based indexes into a single index for ANNS, enabling practical merging for clustering datasets and simplifying index management.
The Semantic Ladder: A Framework for Progressive Formalization of Natural Language Content for Knowledge Graphs and AI Systems introduces the Semantic Ladder, an architectural framework for progressively formalizing natural language data into machine-actionable semantic models, with pattern-based compliance approaches.
I/O Optimizations for Graph-Based Disk-Resident Approximate Nearest Neighbor Search: A Design Space Exploration examines I/O-bound SSD-based ANN and presents an I/O-first framework across memory layout, disk layout, and search algorithm, introducing a page-level complexity model and validating it on four public datasets.