This study proposes an advanced Internet fraud transaction detection method, the Temporal-aware Heterogeneous Graph Oversampling and Attention Fusion Network (THG-OAFN), designed to address the increasingly severe fraud issues in EC. The method innovatively abstracts transaction data into a heterogeneous graph structure, captures temporal dynamic features through Gated Recurrent Unit (GRU), and fuses Graph Neural Network (GNN) to process static topological relationships. To address data imbalance, an improved Graph-based Synthetic Minority Oversampling Technique (GraphSMOTE) framework is introduced, maintaining the structural integrity of fraud clusters through k-hop topological constraints. Meanwhile, a multi-layer attention mechanism (including relationship fusion, neighborhood fusion, and information perception modules) is employed to achieve active fraud prevention. Experimental results show that THG-OAFN attains an area under the curve (AUC) of 96.56% (a 7.78% improvement over the best baseline). Moreover, it achieves a recall of 95.21% (a 6.29% improvement) and an F1-score of 94.72% (a 3.96% improvement) on the Amazon dataset. On the YelpChi dataset, these three metrics reach 90.43%, 89.51%, and 90.31%, respectively, remarkably outperforming existing GNN models. This achievement provides a deployable solution for dynamic fraud detection and active defense. Our code is available at https://github.com/wei4zheng/THG-OAFN.

Scalable multi-task learning for particle collision event reconstruction with heterogeneous graph neural networks

The work is to investigate the trajectory prediction for multiple types of traffic participants in signalized intersection scenarios within intelligent connected environments based on Heterogeneous Spatio-Temporal Multi-Scale Attention Network (HST-MSAN), where participants include Connected and Automated Vehicles (CAVs), Human Vehicles (HVs), cyclists, and pedestrians. A novel method of trajectory prediction that integrates spatio-temporal interaction features and multi-scale map features is proposed based on HST-MSAN. The interaction model is established based on Spatio-Temporal Graph Attention Network (STGAN). The trajectory prediction model is constructed based on STGAN and Multi-Scale Squeeze-and-Excitation Network (MS-SENet). First, an STGAN is developed to differentially encode the historical trajectory of each participant, model the complex interactions, and quantify the interaction intensity among participants. Second, an MS-SENet that integrates Multi-Scale Convolutional (MSC) and a Squeeze-and-Excitation (SE) module is proposed, where multiple parallel convolutional kernels are employed to extract both local and global map features. The proposed model is validated through the INTERACTION dataset. The results of three-second trajectory prediction show that the average displacement error (ADE) and final displacement error (FDE) can reach to 0.17 and 0.68 m, respectively. ADE is reduced by 26.1%, 22.7%, 10.5%, and 29.2%, respectively, and FDE is reduced by 10.5%, 12.8%, 8.1%, and 5.6%, respectively, compared with prediction methods of multiple participants of Heterogeneous Edge-enhanced graph attention network (HEAT), Heterogeneous Driving Graph Transformer (HDGT), Hybrid transformer trajectory network (HTTNet), and Flock-inspired network (FN). The ablation experiments show that ADE is reduced by 22.2% and 19.0%, respectively, and FDE is reduced by 10.0% and 5.6%, respectively, compared with the models without STGAN and without MS-SENet. This model of trajectory prediction jointly models the temporal interaction features of the participants, the spatial interaction features with the surrounding participants, and the multi-scale map features that are most suitable for the current state of the participants. By accurately predicting the future movement trajectories of the surrounding participants, CAVs can identify potential conflict points in advance, optimize their trajectory planning, and reduce the risk of traffic accidents.

Real-world systems frequently exhibit hierarchical multipartite graph structures, yet existing graph neural network (GNN) approaches lack systematic frameworks for hyperparameter optimization in heterogeneous multi-level architectures, limiting their practical applicability. This study proposes a Bayesian optimization framework specifically designed for heterogeneous GNNs operating on three-level graph structures, addressing the computational challenges of configuring partition-aware architecture. Four GNN architectures—Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Graph Isomorphism Networks (GINs), and GraphSAGE—were systematically evaluated using Gaussian Process-based Bayesian hyperparameter optimization with inter-partition message-passing mechanisms. The framework was validated on the TIMSS 2023 dataset (10,000 students, 789 schools, 25 countries), demonstrating that Bayesian-optimized GraphSAGE achieved the highest explained variance (R2 = 0.6187, RMSE = 71.73, MAE = 64.32) compared to seven baseline methods. Bayesian optimization substantially improved model performance, revealing that two-layer architectures optimally capture cross-partition dependencies in three-level structures. GNNExplainer was used to identify the most influential student-level features learned by the model, providing explanatory insight into how the model represents individual characteristics. The optimization framework is broadly applicable to other heterogeneous and multi-level graph settings; however, the empirical findings, such as the optimal architecture depth, are specific to hierarchical graphs with structural properties like the TIMSS topology.

Abstract Predicting tumor sensitivity to therapeutic agents is a central problem in precision oncology, yet developing models that can generalize to new, un-screened cancer types remains a significant challenge. Current precision oncology approaches benefit only a small fraction of cancer patients, partly due to the difficulty of computationally modeling the complex relationships among tumors, somatic mutations, and drug-gene pathways. To address this gap, we present PRELUDE, a heterogeneous graph neural network (GNN) framework designed to leverage these biological relationships to identify cancer cell-specific drug vulnerabilities. Our approach begins with the careful curation of a knowledge graph composed of: (1) drug-cell interactions from large-scale screening panels, (2) drug-gene relationships from curated inhibitory target databases, (3) cell-gene links derived from somatic loss-of-function mutation data, and (4) a comprehensive gene-gene interaction network We show that PRELUDE outperforms existing precision oncology baselines. Our curriculum learning approach forces the model to learn generalizable, biology-driven patterns, demonstrated by its ability to accurately predict responses for cell lines completely removed from the training graph, mimicking the challenge of predicting responses for new patients. Furthermore, our approach is interpretable, identifying effective drug target genes that interact with mutated genes in cancer cells. These findings highlight the potential of graph-based methods to enhance predictive modeling in precision oncology and support their broader adoption in data-driven cancer research. Citation Format: Luis E. Tafoya, Mikaela Dicome, Yue Hu, Macaulay Oladimeji, David Arredondo, Yanfu Zhang, Kushal Virupakshappa, Avinash Sahu. PRELUDE: A graph neural network for drug response prediction [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A033.

Fast Heterogeneous Graph Neural Network Generation via Meta Contrastive Learning.

Predicting drug-target interactions based on multivariate information fusion and graph contrast learning.

In an online search, users often input an ambiguous short query to search engines, which leads to search engines being unable to accurately understand the true users' query intent. Thus, enhancing the users' query intent is necessary. Traditional methods of guessing and inferring user intentions are based on either personal past search data, or the group's search history data. The former faces the cold start problem for new users due to the lack of search history data, while the latter cannot accurately get the intent of new search requests due to different users having different intentions even for the same search query. To solve the above issues and to enhance the representation of search requests by adding some query keywords, we construct a user-query-document search heterogeneous graph with users' search history data of their friend networks, which can express the behavioral features and interrelationships of searches. To facilitate the enhanced representation of a query intent, we present TAHAN, a type-aware heterogeneous graph attention network (GAT) model. Extensive experiments on real-world datasets show that our method not only outperforms the state-of-the-art models, but also achieves superior performance in addressing the data sparsity and cold-start problems.

Background: Long non-coding RNAs (lncRNAs) are a category of more extended RNA strands that lack protein-coding abilities. Although they are not involved in the translation of proteins, studies have shown that they play essential regulatory functions in cells, regulating gene expression and cell biological processes. However, it is both costly and inefficient to determine the associations between lncRNAs and diseases through biological experiments. Therefore, there is an urgent need to develop convenient and fast computational methods to predict lncRNA-disease associations (LDAs) more efficiently. Objective: Predicting disease-associated lncRNAs can help explore the mechanisms of action of lncRNAs in diseases, and this is crucial for early intervention and treatment of diseases. Methods: In this paper, we propose an enhanced heterogeneous graph representation method for predicting LDAs, named GCGALDA. The GCGALDA first obtains the topological structure features of nodes by a biased random walk. Based on this, the neighboring nodes of a node are weighted using the attention mechanism to further mine the semantic association relationships between nodes in the graph data. Then, a graph convolution network (GCN) is used to transfer the neighborhood features of the node to the central node and combine them with the node's features so that the final node representation contains not only structural information but also semantic association information. Finally, the association score between lncRNA and disease is obtained by multilayer perceptron (MLP). Results: As evidenced by the experimental findings, the GCGALDA outperforms other advanced models in terms of prediction accuracy on openly accessible databases. In addition, case studies on several human diseases further confirm the predictive ability of the GCGALDA. Conclusion: In conclusion, the proposed GCGALDA model extracts multi-perspective features, such as topology, semantic association, and node attributes, obtains high-quality heterogeneous graph node representations, and effectively improves the performance of the LDA prediction model.

MAMHSAN: A multi-agent deep reinforcement learning framework based on multi-head self-attention network with heterogeneous graph embedding for flexible job shop scheduling

Generative-Contrastive Heterogeneous Graph Neural Network

A deep reinforcement learning optimization algorithm based on heterogeneous graph neural network for hybrid flow shop scheduling problem with finite transportation resources

Transfer learning with graph neural networks for pressure estimation in monitoring-limited water distribution networks.

Graph neural networks (GNNs) have proven effective in capturing relationships among nodes in a graph. This study introduces a novel perspective by considering a graph as a simplicial complex, encompassing nodes, edges, triangles, and $k$k-simplices, enabling the definition of graph-structured data on any $k$k-simplex. We design a novel Hodge-Laplacian heterogeneous graph attention network (HL-HGAT) to learn heterogeneous signal representations across $k$k-simplices. The HL-HGAT incorporates three key components: HL convolutional filters (HL-filters), simplicial projection (SP), and simplicial attention pooling (SAP) operators, applied to $k$k-simplices. HL-filters leverage the unique topology of $k$k-simplices encoded by the Hodge-Laplacian (HL) operator, operating within the spectral domain of the $k$k-th HL operator. To address computation challenges, we introduce a polynomial approximation for HL-filters, exhibiting spatial localization properties. Additionally, we propose a pooling operator to coarsen $k$k-simplices, combining features through simplicial attention mechanisms of self-attention and cross-attention via transformers and SP operators, capturing topological interconnections across multiple dimensions of simplices. The HL-HGAT is comprehensively evaluated across diverse graph applications, including NP-hard problems, graph multi-label and classification challenges, and graph regression tasks in logistics, computer vision, biology, chemistry, and neuroscience. The results demonstrate the model's efficacy and versatility in handling a wide range of graph-based scenarios.

Abstract At present, graph neural network (GNN) based bearing fault diagnosis faces three major challenges: limited capability in capturing complex multi-scale geometric features of fault graphs, redundant graph connections that hinder effective feature propagation, and insufficient dynamic weighting between nodes, which reduces adaptability under varying operating conditions. To address these challenges, this study proposes a Dual-Stage Heterogeneous Graph Attention Network (DSHGAT) that integrates VGGNet19-based feature extraction with Learned Ollivier–Ricci Curvature (LORC). Specifically, the VGGNet19 module is employed to extract deep features from both time-domain and time-frequency graphs, producing feature matrices that serve as inputs for nodes and edges within the graph structure. Subsequently, topological connections among nodes are optimized through LORC, which eliminates redundant links and enhances graph representation. Finally, node weights are dynamically adjusted at both local and global levels through a dual-stage attention mechanism, thereby enhancing the model’s adaptability to varying operating conditions and improving its overall generalization capability. Experimental results show that the proposed VGGNet19–LORC–DSHGAT model achieves classification accuracies of 99.17% and 99.84% in identifying bearing faults under constant and variable rotational speeds, respectively, thereby confirming the effectiveness of the proposed approach.

EHGFL: Contrastive distillation for efficient heterogeneous graph few-shot learning

Human-associated microbes play a critical role in physiological processes and disease development, including cancer. Predicting microbe-drug associations (MDAs) can aid drug discovery and personalized medicine. However, existing methods cannot predict MDAs involving microbes or drugs absent from labeled data, and they fail to model the underlying biological mechanisms between microbes and drugs. To address these limitations, we propose a novel computational framework, named MetaMDA, for predicting MDAs by performing random walks on a microbe-metabolite-drug heterogeneous network. MetaMDA first constructs a heterogeneous graph that integrates microbes, metabolites, and drugs, enabling the modeling of complex biological interactions. A random walk algorithm with tailored transition probabilities is subsequently applied to the graph, effectively capturing features from multiple node types on a unified scale. Experimental results across multiple datasets demonstrate that MetaMDA consistently outperforms state-of-the-art methods, achieving an average improvement of 26%. Notably, we show MetaMDA's unique ability to predict MDAs involving microbes or drugs absent from labeled data, as illustrated by associations related to acarbose. Furthermore, mechanistic analysis of MetaMDA provides biological explanations for the associations between E. coli and escitalopram, highlighting its potential to reveal a deeper mechanistic understanding of microbe-drug associations. The code and datasets are available on Zenodo https://doi.org/10.5281/zenodo.17348446 and GitHub https://github.com/wqlyt17/MetaMDA. Supplementary data are available at Bioinformatics online.

A heterogeneous graph neural network based entity relationship extraction method in automotive parts supply chain

Federated Heterogeneous Graph Neural Network Enhanced by Perturbation Graph Contrastive Learning and SpectralNet for Privacy-preserving Recommendation

Immune-related adverse events (irAEs) are common and potentially fatal adverse events. However, predicting irAEs based on clinical medication regimens and basic patient information remains a significant clinical challenge. This study aims to develop a prediction model using graph neural networks with electronic health records (EHRs), thereby reducing irAEs risk. Our method based on heterogeneous graph networks. It incorporates medications, diagnoses and patients characteristics from EHRs as nodes to predict irAEs occurrence. Medication-policy simulation, case studies and interpretability analyses were conducted to align the model with real-world clinical needs. Compared to other baseline methods, our method shows superior performance across all evaluation metrics: with AUC of 0.902, AUPRC of 0.85, precision of 0.709, RECALL of 0.799, F1 score of 0.751, accuracy of 0.851. About simulation study, the model demonstrated progressive improvement, reflected in a 5%-6% increase across six evaluation metrics. Interpretability analysis revealed that distinct risk patterns emerge at different treatment stages. Our approach exhibits robust reliability and outperforms other methods for irAEs prediction. Our study further establishes a novel paradigm for personalized therapy monitoring and early intervention. This methodology holds potential for reducing irAEs risk.