Showing 12 papers for 2026-03-16
We propose a defense framework that uses Lyapunov stability to constrain the feature-update dynamics of Graph Neural Networks, covering both integer- and fractional-order stability. This control-theoretic approach aims to improve robustness against adversarial perturbations without relying on heavy adversarial training or data purification, by making the GNN dynamics itself stable and learnable.
Sobolev-Ricci Curvature (SRC) is a graph curvature canonically induced by Sobolev transport geometry, enabling efficient evaluation via a tree-metric Sobolev structure on neighborhood measures. The authors establish two consistency behaviors that anchor SRC to classical transport curvature, providing theoretical grounding and practical utility for graph reweighting and pruning tasks.
We propose an association-aware graph neural network (AAGNN) for learning precoders in cell-free systems. AAGNN explicitly models the dynamic UE-AP association status, leveraging both channel state information and association patterns to learn precoding, and it outperforms association-agnostic baselines.
This work investigates positional encodings for directed graphs, introducing Walk Profile, a generalization of walk-counting sequences that captures numerous structural features important for directed graphs. It provides guidance for designing graph transformers that better encode relative positions in directed networks.
Oversmoothing has been identified as a major limitation of deep GNNs. This work reframes oversmoothing as an optimization issue, examining training dynamics and gradient behavior to explain why extremely deep GNNs still oversmooth despite large weights.
We propose Graph Invariant Learning for Graph Transformers to achieve better out-of-distribution generalization under distribution shifts. The method designs attention mechanisms and positional/structural encodings grounded in invariance principles to capture stable graph patterns across tasks.
CCMamba introduces Topologically-Informed Selective State-Space Networks on Combinatorial Complexes for higher-order graph learning. It reduces the quadratic complexity of existing topological DL methods by selectively modeling state spaces, enabling scalable, rank-aware learning on higher-order structures.
This work examines the geometric coherence of global aggregation in federated GNNs with heterogeneous client graphs. While global aggregation may converge numerically, it can degrade relational behavior; the paper identifies geometric failure modes and proposes remedies to preserve relational integrity.
AutoClimDS is a climate data science agentic AI MVP that integrates a curated climate knowledge graph with agentic AI workflows for cloud-native analysis. The KG unifies datasets, metadata, tools, and workflows to enhance discovery, collaboration, and reproducibility in climate research.
We address miscitation on the scholarly web by combining LLM-augmented text-rich graph learning to reason about citations in context and network structure. Compared with purely semantic or network-based methods, this approach leverages large language models for reasoning about citation context while considering deployment costs.
TRACE is a temporal rule-anchored chain-of-evidence on knowledge graphs for interpretable stock movement prediction. It unifies symbolic priors, dynamic graph exploration, and LLM-guided reasoning in an end-to-end pipeline, performing rule-guided multi-hop exploration and grounding reasoning in contemporaneous news to produce auditable UP/DOWN verdicts with readable evidence paths.
This work formulates transmission topology optimization via busbar splitting as a transferable graph-learning problem, enabling near-real-time solutions for large-scale systems and generalization to unseen topologies and operating conditions across different systems.