Showing 34 papers for 2026-05-08
COPYCOP presents a verification method to determine whether two GNNs were trained independently or whether one was copied and transformed to mimic the other. The adversary can vary architectures, weights, and embedding dimensions, and even transform embeddings to hide the relationship, yet CopyCop can detect such copycat GNNs. The paper shows that CopyCop outperforms existing approaches in identifying copied models.
This work studies the robustness of Graph Self-Supervised Learning (GSSL) when graphs are automatically extracted from text in biomedical domains. It performs a comprehensive evaluation of how real-world noise, not synthetic perturbations, affects GSSL representations and downstream tasks. The findings highlight the need for robust design and benchmarks, and may propose guidelines or datasets.
The authors propose a dual-purpose benchmark that jointly evaluates knowledge graph construction quality and Graph Neural Network performance on downstream tasks. It addresses common issues in automatically built KG, such as noise, fragmentation, and semantic inconsistencies, ensuring that improvements reflect both graph quality and model ability. This benchmark enables fair, interpretable evaluation of KG construction pipelines.
The paper argues that existing adversarial GNN benchmarks use inconsistent experimental settings, hindering fair conclusions. It introduces a practical benchmark protocol for evaluating attacks and defenses under standardized conditions, with datasets, metrics, and baselines to promote reproducibility and fair comparisons.
Full-Spectrum GNN (FSpecGNN) is a second-order generalization of spectral GNNs. It lifts signals from the node domain to the node-pair domain and extends spectral filtering, enabling more expressive and scalable graph representations beyond the 1-WL limit.
CORE (Cross-site OOD Robust brain nEtwork) provides a unified framework for learning brain networks that generalize across unseen sites in fMRI data. It addresses site-conditioned confounds and the limitations of temporally averaged connectivity, enabling robust cross-site inference.
This work investigates how node features interact with graph topology for pooling. It shows that pooling operators perform best when node features are well aligned with the graph structure and formalizes necessary conditions for effective pooling. The results explain why pooling gains vary across datasets.
Federated Cross-Client Subgraph Pattern Detection studies how subgraph patterns that cross partition boundaries become unidentifiable in federated GNNs. It formalizes the problem as structural observability and proposes methods to bridge the representation-equivalence gap across clients while preserving privacy.
Invariant-Based Diagnostics for Graph Benchmarks advocates using permutation-invariant graph invariants as diagnostics to evaluate benchmarks themselves, not just models. The invariants can capture structural heterogeneity and are more expressive than standard GNNs, helping to distinguish benchmark quality from model capability.
On the Safety of Graph Representation Learning introduces GRL-Safety, a multi-axis benchmark for evaluating how graph representations behave under perturbations of signals, contexts, labels, or evidence. It enables systematic safety-focused comparisons of GRL methods beyond clean transfer.
Towards Metric-Faithful Neural Graph Matching develops a theoretical framework that links encoder geometry to the accuracy of neural GED estimation. It analyzes how embedding space structure influences GED quality and provides guidance for encoder design to improve matching performance.
Dynamic Graph with Similarity-Aware Attention Graph Neural Network for Recommender Systems presents a dynamic GNN that uses similarity-aware attention to model evolving user-item interactions and user-user relations. The approach improves recommendation quality by capturing changing preferences and richer relational signals.
Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models proposes using graphlets as discrete building blocks to create a structural vocabulary for KG foundation models. This addresses irregular KG geometry by providing a consistent, shareable representation of local substructures.
Disentangled Generative Graph Representation Learning argues that existing GRL often suffers from entangled representations due to global masking. It advocates disentangled generative modeling to improve robustness and interpretability of graph representations.
ReMAP introduces a neural reparameterization framework for scalable MAP inference in arbitrary-order Markov Random Fields. It optimizes a differentiable relaxation of the MRF energy in an instance-wise fashion, avoiding reliance on supervised data and enabling efficient inference.
Graph Learning Is Suboptimal in Causal Bandits shows that learning the parent set of variables is not always the best strategy for regret minimization under causal structure. Theoretical results reveal scenarios where standard graph learning delays or worsens performance, challenging conventional approaches.
Fed-Listing studies privacy leakage of label distributions in Federated Graph Neural Networks. It introduces a method to infer label distribution from shared updates and demonstrates potential risks for sensitive data in FedGNN deployments.
ChronoSpike proposes an adaptive spiking Graph Neural Network for dynamic graphs, aiming to combine event-driven spiking computation with graph-based temporal modeling. It addresses the high cost of attention-based methods versus recurrent approaches and achieves efficient, context-rich dynamic graph representation.
GraphVec provides cross-domain graph vectorization for graph-level representation learning. It constructs multi-scale global graphs and extracts spectral embeddings to produce transferable fixed-dimensional graph representations across heterogeneous domains without relying on language-models.
Mochi introduces meta-learning to align pre-training and inference for Graph Foundation Models. It critiques reconstruction-based pretraining and demonstrates that meta-learning-based training improves task alignment and efficiency for downstream tasks.
We present a foundation model for event classification in high-energy physics using a Graph Neural Network architecture trained on 120 million simulated proton-proton collision events across 12 processes. The model is pretrained with challenging multiclass and multilabel tasks to learn a robust, general representation of collision data. We evaluate it on seven event-classification tasks, including new physics processes not seen during pretraining and AT experiments.
We address limitations of RL-based schedulers in Kubernetes by proposing AGMARL-DKS, an Adaptive Graph-Enhanced Multi-Agent RL for dynamic scheduling. Our approach uses a graph-structured representation of the cluster and multiple cooperative agents to scale to large heterogeneous clusters, improving stability, resource utilization, and cost efficiency.
We propose real-time actionable modeling of surgical team dynamics using time-expanded interaction graphs, where team members are time-indexed nodes and communication exchanges define directed edges. This representation captures how technical execution and non-technical skills interact over time, enabling real-time analysis and decision support to improve intraoperative performance.
We study interpreting manifolds and embeddings learned by Graph Neural Networks from IoT traffic flows, addressing the opacity of representations in security-critical operations for evolving IoT networks. We analyze the geometric structure of GNN embeddings and propose interpretation methods to reveal the semantic and relational factors driving predictions.
GEM (Graph-Enhanced Mixture-of-Experts) combines language models with graph-structured dialogue understanding and ReAct-style reasoning for dialogue state tracking. The framework dynamically routes between specialized experts—a Graph Neural Network capturing dialogue structure and domain-aware components—to improve DST across domains.
We address heterophily in heterogeneous graphs with HeterSEED, a semantics-structure decoupling framework that separates semantic meaning from structural relations to prevent misleading message passing under heterophily. Experiments show improved performance over strong baselines on standard heterogeneous-graph benchmarks.
We introduce QpiGNN, a quantile-free uncertainty quantification method for GNNs. By directly optimizing prediction-interval coverage and width, it provides reliable uncertainty estimates without resampling or post-hoc calibration typical of quantile regression approaches.
Shadow-Loom provides an experimental framework that turns a narrative into a versioned graphical world model and enables two engines— a causal physics engine grounded in Pearl's causality ladder and a counterfactual calculus over Ancestral Multi-World Networks—to reason about causes, secrets, and consequences in stories. It also scores the same graph against four reader-states—mystery, dramatic irony, suspense, and surprise—bridging causal reasoning with narrative perception.
KGLAMP is a knowledge-graph-guided language-model planning framework for adaptive multi-robot planning and replanning. It uses a knowledge graph to encode robot capabilities and environment, enabling LLMs to produce coordinated plans that account for heterogeneity and dynamic uncertainty, with replanning as conditions change.
DPD-Cancer is an explainable graph-attention deep learning framework for predicting small-molecule anti-cancer activity across the NCI-60 panel under chemistry-aware data partitioning. On hold-out data it achieves AUROC 0.87 and AUPRC 0.73, with accompanying per-cell-line regression results, demonstrating strong predictive performance and interpretability.
Text-Graph Synergy introduces a bidirectional verification and completion framework for Retrieval-Augmented Generation (RAG) to mitigate information island problems. By letting text and graph evidence verify and complete each other in both directions, the method enhances factual grounding and robust multi-hop reasoning.
GATHER tackles zero-shot cell-type annotation with convergence-centric hyper-entity retrieval. For queries composed of dozens to hundreds of expressed genes, it leverages gene co-occurrence patterns to converge on the correct label, addressing limitations of local, entity-by-entity evidence search.
SkipDisk is a disk-memory hybrid ANN search method that reduces memory footprint while maintaining latency comparable to or better than in-memory methods like HNSW. It analyzes existing disk-based approaches and demonstrates low-latency high-dimensional ANN with significantly reduced memory usage.
U-HNSW proposes an efficient graph-based solution for approximate nearest neighbor search under universal Lp metrics, enabling queries for all p in (0,2] without per-p-value indexes. Unlike MLSH, an LSH-based baseline, U-HNSW delivers improved query performance across the universal Lp setting.