graph-learning

DST-Net 提出一种解耦的时空网络用于时间序列预测，针对趋势和季节分量分别设计潜在图块和高斯多尺度注意力机制，以匹配各分量的内在特性并在多个基准上验证有效。

GraphDecisionBench：一个模型无关的基准，用于评估 System One 模型在图结构任务上的决策能力，涵盖 14 种模型-接口配置，揭示了邻接识别准确并不等同于结构正确性等关键发现。

GraphOPD 提出首个将图结构增强引入 on-policy 蒸馏的 agent 训练方法，利用环境状态变化构建步骤依赖图并结合随机游走平稳分布进行信用分配，在 ALFWorld、WebShop、SearchQA 上较最强基线提升最高 5.8 个百分点。

The paper introduces FedCORE, a federated adaptation framework that jointly couples multimodal encoder perception and graph neural network reasoning via a shared low-dimensional latent state, reducing the encoder-GNN pairing gap by 80.7% on decentralized multimodal graphs.

MOVE is a multimodal open-world class verification and expansion framework for graph learning that jointly uses visual tokens, textual attributes, and graph context to detect unknown nodes, generates candidate classes via a multimodal LLM, and selectively expands the class space, achieving an average 11.87% improvement across unknown recognition, open-domain annotation, and downstream tasks.

This paper proposes a moment-guided edge sampling framework that uses spectral moments of the random-walk transition matrix to quantify and control the impact of local edge edits on global graph structure, with applications in graph mining and learning.

This paper proposes HermNet, a spectral graph neural network model using Hermite polynomials, and analyzes its performance in graph filtering tasks through theoretical insights and experiments.

The paper proposes a spectral connectivity-regularized graph learning framework (SCoGL) that incorporates Laplacian spectral priors to improve graph recovery and downstream tasks like graph signal denoising when data is scarce.

The paper introduces a multi-agent agentic graph learning framework (MAAGL) that partitions graphs into communities and uses structural signatures to improve graph reasoning tasks, outperforming state-of-the-art methods.

The paper investigates using graph neural networks to correct spectral prior estimates for connectivity-loss in road networks under disruptions, with evaluations on synthetic and real-world OpenStreetMap datasets.

This paper proposes a graph dictionary learning framework that represents graphs as Gaussian distributions using filtered Laplacians and optimal transport distances, achieving competitive performance in graph clustering and classification tasks.

This paper proposes SLM-Conditioned Hierarchical Relation Routing, an architecture that integrates small language models into graph neural networks to enable adaptive message selection in labeled property graphs, enhancing prediction accuracy by leveraging contextual semantic information.

This paper proposes MS-WDRO, a multi-source Wasserstein distributionally robust graph learning framework that fuses heterogeneous sources via Wasserstein barycenter and minimizes worst-case risk for robust graph topology inference, outperforming baselines in graph recovery and diagnostic utility.

This paper introduces AutoGrable, a method that scores candidate graph constructions for tabular data without training a GNN, using a label-alignment risk based on 1-WL color refinement, enabling cheap search for effective table-to-graph mappings.

This paper proposes PairAlign, a pair-centric graph rewiring framework that uses optimal transport-guided communication alignment to alleviate over-squashing in message-passing neural networks, with theoretical analysis and experiments on standard benchmarks.

This paper introduces NodeJEPA, a joint-embedding predictive architecture for node-level graph self-supervised learning that predicts latent representations of masked structure-aware ego-subgraphs, avoiding reconstruction and hand-crafted augmentations. The method is evaluated on node classification benchmarks and shows competitive performance.

This paper systematically studies hypergraph node classification under label noise, proposing HyperTrust, a robust framework with HyperedgeBoost and HyperedgePrune modules, along with a unified benchmark for evaluating LLN and GLN methods on hypergraphs.

Proposes FedTCR, the first systematic federated multimodal graph learning algorithm that handles task, modality, and topology heterogeneity via topology-aware cross-modal routing and tri-level contrastive learning, outperforming baselines across 7 domains.

This paper studies temporal knowledge graph forecasting under controlled distribution shifts using a synthetic generator that encodes recurrence, homophily, and periodicity. Experiments on seven architectures reveal signal-dependent robustness and limitations in model adaptivity to structural breaks.

The paper proposes Label Influence Propagation (LIP), a model that analyzes and propagates label influences in graph neural networks for multi-label node classification, consistently outperforming state-of-the-art methods.