Showing 35 papers for 2026-05-06
This work presents a methodology to assess bridge importance in urban networks using heterogeneous graph analysis, unsupervised clustering, and automated interpretation with large language models. It tackles open-data-driven multi-dimensional importance quantification and discovers functional archetypes of bridges, enabling interpretable insights for infrastructure resilience.
We propose a stable multimodal graph unlearning method that uses feature-dimension aware quantile selection to edit a graph model selectively. Unlike uniform layer editing, this approach reduces over-editing in high-dimensional cross-modal representations, preserving utility after forgetting. The method aims to enable privacy-preserving, robust forgetting in multimodal GNNs.
CBAM impact on European electricity markets is analyzed with a spatio-temporal GNN framework. The model captures cross-border spillovers and simultaneous changes in electricity prices and carbon intensity. Studied on a subgraph of eight countries, results indicate that CBAM has nontrivial price and CI effects beyond simple price shifts.
GRAFT is a posthoc global explanation framework for GNNs that identifies class-level feature importance profiles. It complements existing motif-based global explainers by attributing input node attributes to predictions, using diversity-guided exemplar selection to cover different decision boundaries.
DP explanations can leak structure; we show that releasing DP-perturbed GNN explanations enables reconstruction of hidden graph structure. Our PRIVX attack exploits Gaussian DP's properties to recover edges. The result highlights privacy risks and motivates stronger defenses for explanations.
We propose adaptive graph-based algorithms for conditional anomaly detection and semi-supervised learning. The approach uses label propagation on a data similarity graph with a fast approximate online solver for the harmonic solution on an approximate graph. By collapsing nearby points into local representatives, distortion is minimized, enabling scalable, adaptive analytics.
GNN-based hierarchy-aware embeddings of knowledge graphs are learned by enriching GNNs with a semantic loss derived from ontologies. The resulting embeddings better reflect domain hierarchies and knowledge. The approach is demonstrated on yeast phenotype prediction and yields box embeddings for KG revisions.
GCSVR framework combines graph convolutional learning with support vector regression to forecast urban air pollution robustly. It captures spatial dependencies among stations and uses SVR to handle nonlinear, nonstationary data and anomalies. The result is robust spatiotemporal forecasts even with anomalous observations.
Graph-Tokenizing LLMs compress graph data into graph tokens and use them as prefixes for querying LLMs. The study systematically evaluates whether these tokens are truly understood by LLMs in the natural-language embedding space, challenging the assumption of full graph-token understanding.
We propose a gauge-invariant graph neural network for Abelian lattice gauge theories, where inputs like Wilson loops enforce local gauge symmetry and message passing preserves it, removing redundant gauge degrees of freedom while retaining expressive power.
Joint relational database generation via graph-conditioned diffusion models. The approach uses graph-conditioned diffusion to jointly generate multi-table relational databases, addressing limitations of single-table generation and autoregressive multi-table methods. It improves parallelism and reduces downstream errors in data synthesis.
When LLM Agents Meet Graph Optimization: An Automated Data Quality Improvement Approach
Full-Graph vs Mini-Batch Training: Comprehensive Analysis from a Batch Size and Fan-Out Size Perspective
Cross-Paradigm Graph Backdoor Attacks with Promptable Subgraph Triggers
New Bounds for Zarankiewicz Numbers via Reinforced LLM Evolutionary Search
Position: How can Graphs Help Large Language Models?
Shadow-Loom: Causal Reasoning over Graphical World Model of Narratives
SCGNN: Semantic Consistency enhanced Graph Neural Network Guided by Granular-ball Computing
Fine-Grained Graph Generation through Latent Mixture Scheduling
SCPRM: A Schema-aware Cumulative Process Reward Model for Knowledge Graph Question Answering
KGQA has advanced with structured query generation, but most work targets RDF/SPARQL, leaving property graphs and Cypher underexplored. UniQGen is a constraint-based framework that uses LLM agents to dynamically extract and refine representative graph query clauses into executable, intent-aligned queries across languages like Cypher. It builds on a variant of Chase & Backchase to guide the clause-level generation.
PhaseNet++ introduces phase-aware anomaly detection in the frequency domain for industrial control systems. By leveraging the phase spectrum from time-frequency transforms and phase coherence graphs, it captures inter-sensor relationships that amplitude-based approaches miss. The approach aims to improve detection of cyber-physical anomalies in ICS beyond traditional amplitude-centric methods.
This survey reviews graph rewiring techniques to mitigate over-squashing and over-smoothing in GNNs, analyzing how topology changes affect information flow. It compares different rewiring methods, discusses their benefits and limitations, and highlights open challenges for future research. The goal is to guide practitioners in selecting rewiring strategies for diverse tasks.
This work studies a multi-base station and multi-RIS pinching-antenna system and jointly optimizes PA placement, RIS phase shifts, transmit beamforming, and BS-UE association to maximize sum rate and energy efficiency. Constraints include inter-PA spacing and power budgets. A GNN-based approach is proposed to solve the problem efficiently with spectral and energy considerations.
Graph-based CFD surrogates like GNNs and Transformers are popular, but training often relies on naive node-wise supervision and explicit Euler time-stepping. This work advocates mesh-based simulations that incorporate spatial and temporal awareness to better capture stiff dynamics and local flux continuity, improving fidelity and efficiency of physics surrogates.
TCDA proposes a thread-constrained discourse-aware modeling for conversational sentiment quadruple analysis in DiaASQ. It addresses structural noise from simple GCNs and distance dilution from standard RoPE by capturing the temporal thread structure and utterance-level progression, enabling more accurate sentiment quadruple extraction.
Causal Software Engineering outlines a vision and roadmap for incorporating causal reasoning into software engineering. It discusses the need to study interventions and counterfactuals in decisions influenced by code, field data, and socio-technical processes, and highlights directions for AI-assisted causal analysis, experimentation, and decision support.
PIEGraph fuses analytical physics with data-driven models to learn object dynamics for rigid and deformable bodies from few interactions. By representing objects as 3D particles and using a neural-equivariant dynamics model, it achieves data-efficient learning and preserves physical feasibility over long horizons.
Intersectional Sycophancy investigates whether LLMs exhibit demographic-conditioned false validation. Across 768 multi-turn conversations with 128 personas varying race, age, gender, and confidence in math, philosophy, and conspiracy domains, the study finds sycophancy varies by model and domain and interacts with perceived demographics.
GraphLand argues that current graph ML benchmarks cover too few domains. It proposes evaluating graph machine learning models on diverse industrial data to assess transferability and generalization of graph foundation models across domains and applications.
Multi-modal Relational Item Representation Learning (MMSC) learns substitutable and complementary items by integrating multiple signals with self-supervision to address weak supervision from behavior data and long-tail sparsity, enabling better recommendations.
Learning Decomposed Contextual Token Representations argues that two-stage generative recommender systems suffer objective misalignment between semantic tokenization and recommender training. It proposes decomposed contextual token representations that jointly leverage pretrained semantic knowledge and collaborative signals.
STAGE presents a full-length screenplay benchmark for reasoning over evolving stories. It combines text, agents, graphs, and evaluation to assess narrative understanding and coherent reasoning across long-form narratives.
Social Bias in LLM-Generated Code examines demographic fairness in code produced by prominent LLMs using the SocialBias-Bench across 343 tasks and seven demographics. The study reveals substantial bias and discusses mitigation strategies.
Deep Interest Mining with Cross-Modal Alignment for SemanticID Generation in Generative Recommendation tackles three limitations of existing Semantic IDs in generative recommender systems: information and semantic degradation and cascaded quantization. It proposes cross-modal alignment and deep interest mining to generate higher-quality SIDs.