graph neural networks

A privacy-preserving federated spatiotemporal dynamic graph neural network framework for epileptic seizure prediction

Abstract Epilepsy ranks among the most prevalent and debilitating neurological disorders, globally affecting an estimated 50 million individuals across all age groups and socioeconomic backgrounds [WHO, 2019]. In recent years, research on its prediction methods has made significant progress, driven by advancements in artificial intelligence technology. Epilepsy prediction models based on electroencephalogram (EEG) signals and deep learning have become an important research direction in the field of neuroscience, and related research results have shown an exponential growth trend. However, there are still several key problems that need to be solved in existing research: First, the mainstream model architecture are focused on traditional neural network framework, exclusively trained in centralized settings that are incompatible with the privacy regulations and data-sharing constraints governing real-world clinical environments; second, the deep learning approaches, including recent Graph Neural Network (GNN) models, using static graph modeling methods, which ignore the dynamic network topological evolution characteristics of EEG signals in the time-varying process, fails to effectively explore the high-order nonlinear correlation characteristics contained in the topological structure of brain functional networks; To address this limitation, this paper proposes a patient-dependent privacy-preserving Federated Learning framework that integrates an epilepsy prediction model DygonNet based on spatiotemporal dynamic graph neural network, a local learning model at each federated client, deployed within a cloud-based simulation environment. In the proposed architecture, the CHB-MIT, SWEC-ETHZ and the TJU-HH iEEG datasets are treated as three independent federated clients representing distinct clinical sites, each performing federated training on their private EEG data, while a central cloud server simulated on Google Colab Pro aggregates the model updates using the Federated Averaging (FedAvg) and FedProx algorithms without ever accessing raw patient recordings. The model defines the dynamic graph structure of EEG signals, innovatively introduces the Transformer model and dynamic graph neural network into the field of epilepsy prediction to fully learn the temporal and spatial characteristics of EEG signals, and proposes a hierarchical graph pooling mechanism based on the attention mechanism in a Federated environment. Experiments show that the model shows excellent epilepsy prediction performance on both public and private datasets.

Yield-graph: multi-stage growth-aware maize yield prediction via graph neural networks

Activated permutation entropy for graph signals

Abstract Nonlinear dynamics play an important role in the analysis of signals. A popular, readily interpretable nonlinear measure is Permutation Entropy (PE). PE has recently been extended for the analysis of graph signals, thus providing a framework for non-linear analysis of data sampled on irregular domains. Integrating ideas from Graph Signal Processing (GSP) and Machine Learning, we introduce Activated Permutation Entropy ($APE_G$) by combining continuous information present in graph signals with traditional ordinal analysis and introducing an ordinal activation function (OAF) akin to the one of neural networks. We also formally extend ordinal contrasts to the graph domain. Activated versions of ordinal contrasts of length 3 are introduced and their advantage is shown in experiments from various domains. Building on recent work in ordinal analysis, we further demonstrate how we can find the continuous ordinal pattern that maximizes the $APE_G$ of various dynamical systems. Simulations with popular non-linear maps and analysis of real-life MRI data show the validity of $APE_G$ and potential benefits over the state of the art. By extending very recent concepts related to permutation entropy to the graph domain, we expect to accelerate the development of improved graph-based entropy methods that enable nonlinear analysis of broader data types and establishing relationships with emerging ideas in data science.

A Fourier physics-informed graph neural network for Allen–Cahn and Conservative Allen–Cahn equations

BioSurveillance-Agent: A Multi-Method Explainable Agentic Platform for Epidemic Surveillance on One-Health Networks

How much does the graph help? An information-parity evaluation of graph neural networks for urban air-quality estimation

Graph neural networks (GNNs) are increasingly used to estimate air pollution at unmonitored urban locations, but reported gains often conflate what a model is with what it sees. We present a heterogeneous, wind-aware graph attention network that embeds regulatory monitoring stations and traffic counters as distinct node types, and evaluate it under an information-parity protocol: six geostatistical and tabular baselines receive the identical, published information set, within a task taxonomy separating virtual sensing (target history available) from spatial extrapolation (no target-side data at any stage). On five stations and 16 traffic counters in Graz, Austria (2018-2021), neighboring stations’ same-day concentrations account for essentially the entire previously observed GNN advantage: the architecture gap at parity is -0.002 R2. In leakage-free extrapolation the GNN does not outperform task-legal geostatistics (-0.06 R2 versus the strongest baseline), a heterogeneous GNNExplainer analysis routes almost all explanation mass through station-to-station rather than traffic edges, and split-conformal intervals under-cover (0.53 and 0.76 at nominal 0.90), quantifying the cost of crossstation calibration. In dense daily-resolution urban networks, evaluation practice rather than architecture dominates reported GNN advantages; the protocol, code, and data are released so the audit can be repeated on any network.

Blockchain fog attention framework for collusion detection and automated accountability in internet of things networks

Fog–edge Internet of Things (IoT) systems support low-latency distributed services but remain vulnerable to coordinated attacks that evade detectors designed for independent events. Existing solutions also tend to separate attack detection, provenance verification, and accountability enforcement, leaving no unified path from relational evidence to auditable response. This study aims to develop an integrated framework that detects coordinated malicious behavior, ranks provenance relevance, verifies evidence, and activates rule-based accountability in resource-constrained edge environments. The proposed Blockchain-Fog Computing Collaborative Framework with Deep Attention-based Collusion Detection and Automated Accountability (BF3-ACDA) framework combines a hierarchical blockchain–fog architecture with an attention-based collusion graph neural network (AttnCol-GNN), whose reputation-aware attention coefficient incorporates behavioral correlation and blockchain-derived trust information. A shared attention representation supports both collusion classification and provenance ranking, while Fog-BFT consensus, Merkle verification, and smart contracts provide tamper-evident recording and severity-based enforcement. Across 30 matched independent runs on the collusion-augmented CICIoT2023 benchmark, BF3-ACDA achieved 96.80 ± 0.23% accuracy, 96.75 ± 0.21% F1-score, and 0.975 ± 0.005 AUC-ROC. Direct-transfer accuracy without target-domain fine-tuning was 94.20 ± 0.34% on NSL-KDD and 92.50 ± 0.25% on CICIDS2017. Removing detector-side reputation and verification features reduced accuracy by 3.40 percentage points, whereas replacing learned attention with mean aggregation reduced it by 1.80 points. The INT8 edge model required 12.4 MB and 58.5 ± 3.9 ms per inference. The results demonstrate a method-level coupling of coordinated-pattern detection, provenance relevance, and auditable enforcement, while supporting prototype deployment on evaluated edge, fog, and cloud platforms. As the collusion metadata were constructed for this study, the findings characterize robustness under controlled conditions rather than field performance on naturally occurring collusion.

Multimodal Graph Neural Networks and Evolutionary Knowledge Fusion for Secure and Explainable Governance in Intelligent IoT Interactive Media Systems

This research proposes a secure, explainable, and context-aware governance framework for blockchain-based digital media contracts in multimodal artificial intelligence–enabled AIoT interactive systems. As digital licensing, NFT copyright management, royalty distribution, and cross-chain content circulation become increasingly embedded in smart media ecosystems, existing contract auditing approaches remain limited by unimodal analysis, weak explainability, black-box decision processes, and insufficient cross-platform generalization. To address these challenges, the proposed framework integrates cross-modal data fusion, graph neural networks, and evolutionary expert knowledge fusion to model smart contract code, abstract syntax trees, control-flow graphs, intercontract transactions, cross-chain records, licensing metadata, NFT copyright attributes, and multimodal IoT interaction logs as unified graph representations. By incorporating causal priors and differentiable rules, the framework supports transparent risk reasoning, verifiable explanations, and trustworthy decision support. The study contributes to explainable blockchain security, multimodal AI governance, and intelligent media systems by enabling more robust detection of copyright misuse, unauthorized licensing, abnormal content distribution, and cross-chain transaction risks.

Solar Photovoltaic Generation Forecasting: A Review of Artificial Intelligence Approaches

The rapid global expansion of solar photovoltaic (PV) capacity has increased the operational need for accurate generation forecasting to support grid balancing, dispatch, and market participation. Artificial intelligence (AI) and machine learning (ML) methods now dominate this research area, but the resulting literature is large and methodologically fragmented, making it difficult to establish which methods are used, what data they require, and where the principal gaps lie. This paper combines a bibliometric analysis of 3111 records retrieved from the Web of Science Core Collection (2010–2026) with a technical synthesis of 27 highly cited studies published from 2022 onward, combining the most highly cited works with targeted additions from 2024–2025 covering specific methodological gaps. The bibliometric analysis shows exponential growth in annual output, from three publications in 2010 to 609 in 2025, with keyword evolution tracing a clear methodological trajectory from classical and fuzzy-logic approaches, through shallow and deep neural networks, to transformer- and attention-based architectures since 2023. The technical synthesis finds that classical machine learning remains competitive for day-ahead forecasting with well-structured numerical weather prediction inputs, that convolutional neural network–long short-term memory (CNN-LSTM) hybrids dominate the deep-learning literature, and that graph-based and transformer architectures address multi-site and multi-horizon forecasting, respectively. A comparison of reported results shows that absolute error metrics are not directly comparable across studies due to heterogeneous datasets, metrics, temporal resolutions, and climates, although relative improvements within controlled comparisons are directionally consistent. Seven research gaps are identified, including the absence of standardized benchmarks, limited public dataset availability, weak cross-region generalization, and underdeveloped uncertainty quantification.

Learning Subgroup Relations Using Siamese Graph Neural Networks

Determining whether one finite group is isomorphic to a subgroup of another is a fundamental problem in computational group theory. In this work, we propose a Siamese Graph Neural Network (Siamese GNN) for subgroup prediction using Cayley graph representations of finite groups. Each input group is represented by its undirected Cayley graph and encoded by one branch of a Siamese GNN to produce a graph embedding. The resulting graph embeddings are combined with algebraic features derived directly from the input groups to construct a joint feature vector, which is processed by a fully connected classifier to predict subgroup relations between finite groups. By integrating graph-based structural representations with algebraic features, the proposed framework provides a unified approach for learning subgroup relations from finite groups. Experimental results on an expanded and more diverse dataset of 308 finite-group pairs drawn from 11 group families demonstrate the effectiveness of the proposed architecture, achieving a test BA of 91.67% on an independent test set. Additional experiments evaluate generalization to unseen groups, robustness to different Cayley graph generating sets, the contribution of GNN message passing, performance relative to non-neural baselines, and comparison with exact computational methods. These results illustrate the potential of geometric deep learning for subgroup prediction.

A Parallel Convolutional–Hypergraph Feature learning Network for pulmonary nodule classification

Revisiting topological bias in imbalanced graph learning: A Dual-Balanced Graph Neural Network approach