Showing 27 papers for 2026-05-11
Model-to-Data (M2D) distillation transfers model complexity from GNNs into the data representation, increasing transparency of graph learning. By shifting complexity to the data, M2D clarifies the performance gap between simple and more powerful models and provides a more interpretable framework than traditional explainability methods.
We introduce a GRU-gated Graph Attention Network for inductive transfer in power-grid cascading failures. Trained on combined data from limited grids and applied directly to unseen grids without retraining. The model identifies vulnerable transmission lines before cascades.
We propose a Max-Cut specific graph neural network that serves as a feasibility-preserving proxy for SDP relaxations used within a branch-and-bound framework. This lightweight neural solver plugs into exact B&B to obtain global-optimal Max-Cut solutions more efficiently.
We present a multi-dimensional benchmark for Graph Anomaly Detection (GAD) that tests models under deployment-relevant challenges such as million-scale graphs and extreme anomaly ratios. The benchmark aims to close the gap between academic evaluation and real-world deployment.
We propose learning multi-relational graph representations for DNA methylation-based biological age estimation, capturing complex relationships among CpG sites. This relational approach improves age predictions from methylation biomarkers.
We analyze why self-inconsistency arises in self-interpretable GNN explanations. We identify re-explanation-induced context perturbation as the direct cause and propose a latent signal assignment hypothesis to explain edge-sensitive variations.
We revisit bilevel graph structure learning and show that inner-loop training dynamics explain a large part of the reported gains, not just graph rewiring. We introduce frozen-phi to freeze the graph while keeping the inner-loop schedule, isolating the contribution of rewiring vs dynamics.
We propose Graph Representation Learning Augmented Model Manipulation for Federated Fine-Tuning of LLMs (FFT). The method detects manipulated updates and robustly aggregates them to defend against adversarial participants in privacy-preserving federated adaptation.
GRAPHLCP is a structure-aware localized conformal prediction framework on graphs. It uses proximity in graph structure to localize conformal predictions and provide finite-sample guarantees, addressing embedding-based localization limitations.
We propose GenTTP, a generalized travel time predictor that accounts for varying route choices in urban networks. It learns to uncover complex spatial patterns and predict flow and travel time under multiple demand realizations.
We propose cross-atlas consistent brain disorder representations via disentangled multi-atlas functional connectivity learning. By disentangling atlas-induced variations, the method yields more consistent representations across atlases for brain disorder analysis.
AdaTKG introduces adaptive memory for temporal knowledge graph reasoning. Each entity has an adaptive memory that is refined whenever the entity participates in events, enabling more accurate temporal reasoning.
Topic Is Not Agenda conducts a citation-community audit of text embeddings. Using an augmented citation graph and Leiden CPM clustering, embeddings align with broad sub-fields (L1) but fail to capture research agendas (L2), with only about 15% top-10 same-rate.
GESR introduces graph-based edge semantic reconstruction for stealthy communication detection with benign-only training. It captures relational context in flow graphs to detect covert malicious communications without attack labels.
GRASP is a graph-based anomaly detector using self-supervised classification on provenance graphs for intrusion detection. It reduces dependence on predefined thresholds and improves robustness by leveraging system provenance.
PPI-Net is a hierarchical graph neural network that integrates protein-protein interaction networks with pathway-level representations to model disease from molecular interactions to functional processes, enabling patient-specific predictions.
HYPER is a foundation model for inductive link prediction with knowledge hypergraphs, capable of generalizing to unseen entities and unseen relation types in dynamic relation vocabularies.
Flock is a knowledge graph foundation model learned on random walks to enable zero-shot link prediction. By relaxing strict equivariance, Flock increases expressiveness and transferability across similarly structured KGs.
We study the robustness of spatio-temporal graph neural networks for fault location in partially observable distribution grids. The work analyzes topology choices and measurement sparsity, offering guidelines for robust fault localization.
We present a hybrid graph neural network for EEG-based depression detection that combines fixed brain connectivity graphs to capture common patterns with adaptive connections for individual differences, improving detection performance.
GraphReAct extends the idea of reasoning-acting LLM frameworks to graph inference. It enables interleaved reasoning and actions to fetch informative evidence from graphs and progressively refine context over multiple steps, addressing the structured nature of graphs. The approach leverages graph topology and latent representations to improve multi-step inference.
An embarrassingly simple graph heuristic reveals that many sequential recommendation benchmarks can be solved with minimal modeling. The study audits benchmarks by starting from the last one or two interacted items and retrieving candidates from a few-hop neighborhood using a simple graph rule. The findings challenge the necessity of complex generative models for these benchmarks.
DCGL proposes Dual-Channel Graph Learning that combines knowledge-graph aware signals with language-model embeddings for knowledge-aware recommendation. It addresses implicit semantic relationships beyond explicit KG links and the suboptimal fusion of ID and LLM embeddings, which can cause signal interference. The dual-channel design aims to integrate structured KG information with LLM-derived semantics to boost recommendations.
Response-G1 introduces explicit scene-graph modeling for proactive streaming video understanding with Video-LLMs. It aligns accumulated video evidence with the query's response conditions through structured scene graphs, enabling timely and relevant responses. The approach covers three stages that are fine-tuning-free: online query-guided scene graph generation, memory-based reasoning, and response generation.
Federated Spatiotemporal Graph Learning for Passive Attack Detection in Smart Grids proposes a graph-centric, multimodal detector that works across distributed grid nodes to identify passive eavesdropping. It fuses spatiotemporal signals from multiple sources to reveal faint, short-lived patterns that are hard to detect at a single node or timeline. The framework uses federated learning to protect data privacy while leveraging graph-based modeling to improve detection performance.
TCDA proposes Thread-Constrained Discourse-Aware Modeling for Conversational Sentiment Quadruple Analysis. It addresses limitations of standard GCNs that introduce structural noise and RoPE-based methods that fail to separate token-level order from utterance-level progression and suffer distance dilution. The method constrains modeling to dialogue threads to better capture discourse and sentiment relations across turns.
Towards a theory of Façade-X data access: satisfiability of SPARQL basic graph patterns analyzes when SPARQL basic graph patterns are satisfiable when querying data sources accessed via Façade-X. Façade-X provides a meta-model that specializes RDF to fit diverse data formats, enabling SPARQL queries over heterogeneous sources. The paper develops a theoretical framework characterizing satisfiability and explores expressive limits and practical implications.