A new arXiv preprint examines recurrent graph neural networks, which repeatedly run message passing until it converges. The authors note that existing logical characterizations of these models rely on multi-set aggregation, counting-based graded logics, and halting or acceptance conditions that cannot be verified from the network's parameters.
A new arXiv preprint addresses node classification on evolving graphs, where classifiers must generalize to newly arriving nodes despite distribution shift while also providing calibrated uncertainty. The authors propose an online Bayesian approach aimed at inductive settings where safety-sensitive applications require trustworthy confidence estimates. The work targets a gap left by standard graph neural networks, which typically assume static graphs and offer limited uncertainty quantification.
A revised arXiv paper proposes a graph neural network framework for diagnosing which specific sensor or component is responsible for anomalies in multivariate industrial time series. The work aims to go beyond simple anomaly detection by making the source of the fault interpretable. It targets reliability and safety monitoring in complex, multi-sensor process environments.
A new arXiv paper investigates how well Graph Transformers generalize when trained on graphs of one size and evaluated on larger or smaller ones. It focuses on positional encodings derived from graph neural networks, a common design element in these attention-based models for graph data. The work examines whether such encodings help preserve performance as graph size changes.
A new arXiv preprint introduces a graph-guided mixture-of-experts approach for predicting tumor survival outcomes using multiple data modalities. The method aims to combine heterogeneous clinical inputs within a single predictive framework. The paper appears in the cs.LG category as a new submission.
Researchers introduce LiftGCN, a graph learning method designed to predict finite element stress fields that contain sharp gradients near holes, notches and load points. The approach uses a Joukowski spectral lifting transform to preserve energy and retain high-frequency graph components that standard graph neural networks tend to smooth away. The work is posted as an arXiv preprint in computer science categories.
A new arXiv paper proposes GNN4PPM, which applies relational graph convolutional networks to predictive process monitoring. The method targets several predictions at once, such as the next event in a running process, the time remaining until a trace finishes, and its eventual outcome. The authors argue that existing techniques typically address only one of these targets; the posted abstract is truncated before any experimental results are described.
A new arXiv paper proposes using geometric flow techniques to enhance graph coarsening, a pooling step used in graph convolutional networks to cut computational cost. The approach aims to reduce the expense of graph pooling operations that mirror pooling in standard convolutional networks. The work is a preprint and reports on method design rather than deployed results.
A new arXiv preprint proposes deformable graph neural networks built on positional encoding to tackle long-standing limits of message passing over first-order neighbors. The authors note that conventional GNNs struggle as depth increases, leading to over-smoothing and difficulty capturing long-range structure. The work targets graph-structured data in real-world settings where deeper models are needed.
A new arXiv preprint introduces an instance-aware graph modeling approach for detecting and classifying cells in pathology images. The method aims to capture complex cellular interactions within the tumor microenvironment rather than relying only on visual appearance. Accurate cell detection matters for diagnostic accuracy and treatment planning.
A new arXiv paper addresses a limitation in graph neural networks used for spatio-temporal traffic prediction, which often rely only on sensor proximity or road-network topology. The authors propose a method that injects general semantic knowledge into the model to improve forecasting accuracy. The work is a revised submission (v2) to arXiv's machine learning category.
A new arXiv paper proposes an unsupervised graph neural network approach to the Minimum Dominating Set problem, an NP-hard combinatorial task. The method is aimed at applications such as influence maximization in social networks, viral marketing, and public health interventions. The work frames dominating set selection as a learning problem that does not require labeled optimal solutions.
A new arXiv paper proposes a molecular representation learning approach that combines multiple views rather than relying on a single one. It pairs hierarchical graph modeling of atom-bond topology with contextualized molecular fingerprints to improve property prediction. The goal is representations that generalize from limited labeled data to structurally novel compounds.
A revised arXiv paper argues that neural algorithmic reasoning, which typically trains networks to imitate classic algorithms via supervised learning, is limited by its reliance on post-processing to produce valid outputs. The authors propose reformulating the task as a reinforcement learning problem, aiming to let models build correct solutions directly rather than repairing them afterward. The work appears in the cs.LG and cs.AI listings as a replacement submission.
A new arXiv paper introduces HiGFRL, a method that fuses hierarchical graph representations with reinforcement learning to schedule tasks with dependencies across heterogeneous cloud clusters. The authors target the difficulty of jointly handling DAG structure and multi-dimensional resource limits in online settings, which they say existing deep RL schedulers address only partially. The work is a preprint and has not yet been peer reviewed.
A revised arXiv preprint presents a graph neural network approach for predicting links in temporal signed networks, which capture how cooperative and adversarial relationships evolve over time. The authors motivate the work with applications including social media analysis, trust and reputation systems, and financial transaction networks. The paper's abstract excerpt focuses on the method's design for handling dynamic, sign-aware graph structure.
A new arXiv preprint introduces HGTO, a unified graph-based, physics-informed framework for density-based structural topology optimization. The approach reframes the usual nested loop of material updates, structural analysis, and sensitivity computation, drawing on neural density parameterization and dual-field physics-informed methods that require no labeled data. The abstract positions the work as a data-free alternative to conventional optimization pipelines.
A new arXiv paper argues that existing counterfactual explanation methods for temporal graphs only show how to alter past events to flip a prediction, without stating what the resulting outcome should become. The authors introduce "specified-foil" counterfactuals, which target a particular alternative outcome rather than leaving the replacement undefined. The work aims to give users more actionable explanations of predicted events in time-evolving graph data.
A new arXiv paper proposes moving the transfer mechanism used by knowledge graph foundation models out of specialized architectures and into the data representation itself. The authors treat reification as a transferable vocabulary, allowing plain graph neural networks to perform zero-shot link prediction on previously unseen knowledge graphs. This approach aims to match dedicated models such as ULTRA without requiring architecture-level hard-coding of transfer behavior.
Researchers present a dual-channel graph neural architecture that predicts which exact solver will perform best on a given maximum clique problem instance. Since no single solver dominates across all types of graphs, the approach learns from graph characteristics to make per-instance algorithm choices. The paper appears on arXiv in both the AI and machine learning categories.
A new arXiv paper presents an efficient graph neural network method for optimizing multicarrier wideband hybrid beamforming, a key technique for highly directional 6G links. The work targets beam squint, a distortion that grows as 6G systems use much wider frequency bands and that traditionally requires costly true-time-delay filters. The authors position the learning-based approach as a more efficient alternative for next-generation wireless systems.
A newly published feature article lays out the mathematical underpinnings of covariance neural networks, a class of graph neural networks that treat covariance matrices as graph structures. The work connects classical dimensionality-reduction techniques such as PCA with modern graph-based learning, and highlights the broad range of domains where covariance data naturally occurs.
An arXiv paper pairs graph neural network-guided coarsening with adaptive penalty tuning to shrink the QUBO formulations that arise when capacitated vehicle routing problems with time windows are solved on a quantum annealer. Customers that are geographically close and have compatible delivery windows are merged into super-nodes, reducing the number of binary variables before the problem reaches the annealer. The goal is to make quantum annealing tractable for larger, more realistic logistics instances.
A new arXiv paper proposes a defense for graph neural networks that strips out attacker-inserted edges from a graph's topology before inference, requiring no retraining of the model. The method relies on kernel-complexity signals to identify edges likely introduced through adversarial manipulation, and the authors support the design with theoretical guarantees. It addresses threats that target graph structure rather than node features.
A revised preprint examines how the modular structure of relational graphs interacts with the depth of neural networks when learning from graph-structured data. The author situates the work within graph-based machine learning, including graph neural networks and reinforcement learning, and analyzes how graph structure shapes learning outcomes. The version posted is an update to an earlier draft.