Neurocomputing最新文献

{"title":"Momentum gradient-based untargeted poisoning attack on hypergraph neural networks","authors":"Yang Chen ,&nbsp;Stjepan Picek ,&nbsp;Zhonglin Ye ,&nbsp;Zhaoyang Wang ,&nbsp;Haixing Zhao","doi":"10.1016/j.neucom.2025.129835","DOIUrl":"10.1016/j.neucom.2025.129835","url":null,"abstract":"<div><div>Hypergraph Neural Networks (HGNNs) have been successfully applied in various hypergraph-related tasks due to their excellent higher-order representation capabilities. Unfortunately, recent works have shown deep learning models vulnerable to diverse attacks. Most studies of attacks on graphs have focused on Graph Neural Networks (GNNs), and the study of attacks on HGNNs remains largely unexplored. In this paper, we try to bridge this gap. We design a new untargeted poisoning attack for HGNNs, MGHGA, which focuses on modifying node features. We consider the process of HGNNs training and use a surrogate model to implement the attack before hypergraph modeling. Precisely, MGHGA consists of two parts: feature selection and feature modification. We use a momentum gradient mechanism to choose the attack node features in the feature selection module. In the feature modification module, we use two feature generation approaches (direct modification and sign gradient) to enable MGHGA to be employed on discrete and continuous datasets. We conduct extensive experiments on seven benchmark datasets to validate the attack performance of MGHGA in the node and the visual object classification tasks. The results show that MGHGA improves performance by an average of 2% compared to the baselines.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129835"},"PeriodicalIF":5.5,"publicationDate":"2025-03-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143561945","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"LN-DETR: An efficient Transformer architecture for lung nodule detection with multi-scale feature fusion","authors":"Jiade Tang ,&nbsp;Xiao Chen ,&nbsp;Linyuan Fan ,&nbsp;Zhenliang Zhu ,&nbsp;Chen Huang","doi":"10.1016/j.neucom.2025.129827","DOIUrl":"10.1016/j.neucom.2025.129827","url":null,"abstract":"<div><div>Lung cancer remains one of the leading causes of mortality worldwide, and early detection is crucial for improving patient survival rates. Traditional lung nodule detection methods are inefficient and inaccurate, making them inadequate for clinical needs. Although deep learning methods have made progress in medical image analysis, existing approaches still perform poorly in detecting small, morphologically complex lung nodules, leading to missed detections and false positives. Additionally, the high computational complexity of previous models hinders real-time detection. To address these challenges, this study proposes a Transformer-based lung nodule detection model called LN-DETR. The model integrates a Partial Convolution-based Efficient Multi-scale Attention (PC-EMA) module, a Grouped and Shuffled Convolutional Cross-scale Feature Fusion (GS-CCFM) module, and introduces a Channel Transformer (CTrans) module. PC-EMA combines Efficient Multi-Scale Attention with partial convolution to enhance multi-scale feature extraction while optimizing computational efficiency. GS-CCFM uses Grouped and Shuffled Convolution (GSConv) to achieve efficient cross-scale feature fusion. The CTrans module employs a cross-channel attention mechanism to further strengthen feature fusion capabilities. Experimental results on the LUNA16 and Tianchi lung nodule datasets demonstrate that LN-DETR outperforms existing object detection models in detection accuracy, computational efficiency, and model complexity. On the LUNA16 dataset, LN-DETR achieved an F1 score of 91.5% and a mean Average Precision (mAP) of 93.1%; on the Tianchi dataset, the F1 score was 87.4% and the mAP was 86.4%, both significantly higher than baseline models. Furthermore, the reduced parameter count and computational overhead make the model more suitable for broader clinical applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"633 ","pages":"Article 129827"},"PeriodicalIF":5.5,"publicationDate":"2025-03-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548890","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Memo-UNet: Leveraging historical information for enhanced wave height prediction","authors":"Teng Fang ,&nbsp;Xiaojie Li ,&nbsp;Canghong Shi ,&nbsp;Xian Zhang ,&nbsp;Wei Xiao ,&nbsp;Yi Kou ,&nbsp;Imran Mumtaz ,&nbsp;Zhan ao Huang","doi":"10.1016/j.neucom.2025.129840","DOIUrl":"10.1016/j.neucom.2025.129840","url":null,"abstract":"<div><div>Wave height prediction is a challenging spatiotemporal modeling task. It requires accurately capturing the dynamic evolution patterns of historical data. However, current methods show significant limitations in storing, retrieving, and utilizing historical information. These limitations hinder the learning ability of neural networks in critical temporal dynamics utilization of wave height prediction. Therefore, we propose a novel neural network Meme-Unet for wave height prediction. Specifically, to address the challenges in storing and retrieving historical information, we design a Memo module that adaptively stores historical data and retrieves key historical information through a minimum distance constraint. Additionally, to enhance the utilization effectiveness of historical information, we designed a TimeEncoding module with attention mechanisms to guide the model in better capturing temporal relationships. We conducted wave height prediction experiments in two different ocean regions, achieving the best MSE of 0.041 and a 6.8% improvement compared to state-of-the-art methods. This shows the advantage of the Memo-UNet’s design in capturing the dynamics of spatiotemporal modeling.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129840"},"PeriodicalIF":5.5,"publicationDate":"2025-03-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548124","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DiffuseVAE++: Mitigating training-sampling mismatch based on additional noise for higher fidelity image generation","authors":"Xiaobao Yang ,&nbsp;Wei Luo ,&nbsp;Hailong Ning ,&nbsp;Guorui Zhang ,&nbsp;Wei Sun ,&nbsp;Sugang Ma","doi":"10.1016/j.neucom.2025.129814","DOIUrl":"10.1016/j.neucom.2025.129814","url":null,"abstract":"<div><div>Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated remarkable results in image generation. However, there exist a mismatch between the training and sampling process in current diffusion models, in addition, the U-Net denoising network based on simple residual blocks cannot predict noise information accurately, which affects the generated quality. To address these limitations, we present a novel image generation method that achieves higher fidelity. First, by additionally adding the standard Gaussian noise in the diffusion forward process, which does not disrupt the forward process, our method alleviates the mismatch. Subsequently, an important efficient denoising network based on U-Net is presented, where our proposed Simple Squeeze-Excitation and Simple GLU, combined with Depthwise Separable Convolution, enhance the ability of the model to predict real noise using the Simplified Nonlinear No Activation (SNNA) block. Furthermore, considering the structural characteristics of the baseline model, we introduce an additional cross-attention mechanism to enable DDPM to focus on VAE stage characteristics. Allowing the model to more accurately capture and learn the noise information. Finally, it is shown after extensive experiments the proposed DiffuseVAE++ obtains significant gains in FID scores, improving from 3.84 to 2.41 on CIFAR-10 and from 3.94 to 2.30 on CelebA-64. In particular, the IS scores on CIFAR-10 reaches 10.10, which is comparable to the current state-of-the-art methods competitively (<em>e.g., U-ViT</em>, <em>StyleGAN2</em>).</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"633 ","pages":"Article 129814"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548891","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GKA: Graph-guided knowledge association for fine-grained visual categorization","authors":"Yuetian Wang ,&nbsp;Shuo Ye ,&nbsp;Wenjin Hou ,&nbsp;Duanquan Xu ,&nbsp;Xinge You","doi":"10.1016/j.neucom.2025.129819","DOIUrl":"10.1016/j.neucom.2025.129819","url":null,"abstract":"<div><div>Fine-grained visual categorization aims to distinguish highly similar subclasses by exploiting subtle differences. However, existing methods predominantly emphasize the extraction of visual cues from individual images, overlooking the exploration of semantic relationships within and across classes. To this end, we introduce a novel approach termed Graph-based Knowledge Association (GKA). Specifically, we employ a positional embedding to model the relationship between instances in the feature space, and adaptively mine the connections between features of different instances via a graph neural network. The framework effectively aggregates features from neighboring nodes to enhance the understanding of discriminative features by exploiting complementary information between instances. Furthermore, a plain knowledge-guided module embeds this relational knowledge into the training of the backbone network for discriminative feature extraction, thus improving fine-grained classification performance. Empirical evaluations on four benchmark datasets for Fine-grained Visual Categorization (FGVC) demonstrate that our method achieves state-of-the-art (SOTA) performance.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129819"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548123","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A review of SCADA-based condition monitoring for wind turbines via artificial neural networks","authors":"Li Sheng ,&nbsp;Chunyu Li ,&nbsp;Ming Gao ,&nbsp;Xiaopeng Xi ,&nbsp;Donghua Zhou","doi":"10.1016/j.neucom.2025.129830","DOIUrl":"10.1016/j.neucom.2025.129830","url":null,"abstract":"<div><div>In recent years, the supervisory control and data acquisition (SCADA) data has gained increasing research attention in the field of wind turbine condition monitoring. Artificial intelligence (AI) techniques have been widely applied to address condition monitoring challenges, and artificial neural networks (ANNs), recognized as a foundational component of modern AI, have proven to be particularly effective tools. Wind turbine condition monitoring focuses on analyzing the operational parameters of turbines to realize early fault detection, precise diagnostics, and accurate prognostics, thereby mitigating the risk of catastrophic faults, enhancing system reliability, and improving wind farm operational efficiency. Due to inherent issues in raw SCADA data, including missing values and abnormal data, preprocessing steps such as data cleaning are critical before feeding the data into ANN models. Additionally, the choice of ANN architecture typically depends on the specific requirements of condition monitoring tasks (e.g., fault detection, diagnosis, or prediction/prognosis) and the characteristics of SCADA datasets such as imbalance problem of fault samples. Hence, current research with respect to wind turbine condition monitoring generally follows two approaches: (1) utilizing classification models to identify fault types at specific time points, and (2) employing regression models to construct normal behavior models (NBMs) or track and predict continuous key performance indicators. This survey systematically reviews SCADA-based wind turbine condition monitoring methods within five years, emphasizing neural networks as key approaches, and structures the discussion around three core aspects: data preprocessing, classification models, and regression models. Moreover, the comparative strengths, capabilities, and limitations of various ANNs in each link are discussed. By providing an in-depth analysis, this paper aims to offer theoretical and practical insights to support the further development of condition monitoring technologies for wind turbines.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"633 ","pages":"Article 129830"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548958","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Advancing DoA assessment through federated learning: A one-shot pseudo data approach","authors":"Thomas Schmierer,&nbsp;Tianning Li,&nbsp;Di Wu,&nbsp;Yan Li","doi":"10.1016/j.neucom.2025.129812","DOIUrl":"10.1016/j.neucom.2025.129812","url":null,"abstract":"<div><div>Accurately measuring the Depth of Anaesthesia (DoA) during surgical procedures is crucial for patient safety. A significant challenge in developing effective machine learning models for DoA assessment is the lack of data from single organisations and preserving data privacy between institutions. Federated learning offers a solution by enabling multiple parties to collaboratively train models without exchanging data. However, traditional federated learning algorithms perform poorly in data heterogeneous, non-identically distributed data distribution scenarios. To address these challenges, we propose a one-shot federated learning framework, DoAFedP-NN, which facilitates federated learning with heterogeneous model development. The framework is tested in a range of model and data heterogeneity environments. This method enables the training of a global DoA prediction model across different medical facilities without sharing local data.</div><div>The DoAFedP-NN model, utilising neural network design with entropy and spectral feature extraction, is compared to benchmark federated learning architectures, demonstrating its advantage in handling heterogeneous medical data. Experimental results show that DoAFedP-NN achieves robust DoA estimation when compared to the Bispectral (BIS) index, with high correlation coefficients of 0.8472 and 0.8542 across independent databases. The proposed model outperforms locally developed models, showing significant improvements when validated against external datasets from different medical facilities. This paper makes the key contributions: (1) introduces a one-shot pseudo-data method for federated learning; (2) demonstrates the effectiveness of this approach for EEG-based DoA using real-world databases; (3) showcases the model’s ability to achieve high correlation with the BIS index while preserving patient privacy in a range of client distribution scenarios and under cross-validation.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129812"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548143","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Local and Global Spatial–Temporal Transformer for skeleton-based action recognition","authors":"Ruyi Liu,&nbsp;Yu Chen,&nbsp;Feiyu Gai,&nbsp;Yi Liu,&nbsp;Qiguang Miao,&nbsp;Shuai Wu","doi":"10.1016/j.neucom.2025.129820","DOIUrl":"10.1016/j.neucom.2025.129820","url":null,"abstract":"<div><div>Skeleton-based action recognition represents a dynamic and expanding research domain in computer vision. Currently, GCN-based methods primarily rely on the graph topology to capture dependencies between joints, however, they are limited in the ability to capture long-distance dependencies. On the other hand, transformer-based methods have also been applied to skeleton-based action recognition, as transformers prove effective in capturing long-distance dependencies. Nevertheless, most transformer-based methods directly calculate pairwise global self-attention of all nodes in both spatial and temporal dimensions, making it challenging to distinguish the correlation between short-distance joints and underestimate the impact of short-term temporal dynamics. Additionally, some existing methods often utilize multi-stream fusion to combine features from different modalities, neglecting the fusion of low-level information from these modalities. In this work, we propose a novel <strong>L</strong>ocal and <strong>G</strong>lobal <strong>S</strong>patial–<strong>T</strong>emporal Trans<strong>former</strong>network (LG-STFormer) containing two key components: (1) LGA-module: local and global attention module. The LGA-module enables the model to capture richer temporal and spatial information. It consists of two parts: skeleton topology constraint spatial transformer (STC-SFormer) and attention-enhanced multiscale TCN (AM-TCN). The STC-SFormer focuses on the correlation between local joints and distant joints in the spatial dimension, while the AM-TCN integrates the global self-attention mechanism into multi-scale temporal convolution to capture local and global temporal motion patterns of joints effectively. (2) JBC-Fusion-module: The module consists of joint-bone cross fusion transformer (JBC-Former) and AM-TCN. Utilizing dynamic generation method for complementary features, the JBC-Former facilitates the fusion of low-level information between complementary features. Finally, we make extensive experiments on the NTU-RGB+D, NTU-RGB+D 120, and NW-UCLA datasets to show the competitive performance of the proposed LG-STFormer in the field of skeleton-based action recognition.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129820"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143562522","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"HCKGL: Hyperbolic collaborative knowledge graph learning for recommendation","authors":"Huijuan Hu ,&nbsp;Chaobo He ,&nbsp;Xinran Chen ,&nbsp;Quanlong Guan","doi":"10.1016/j.neucom.2025.129808","DOIUrl":"10.1016/j.neucom.2025.129808","url":null,"abstract":"<div><div>Recently, the integration of knowledge graph and recommendation system has become a hot topic. Its popular solution is firstly combining the knowledge graph and user–item interaction graph to generate a unified Collaborative Knowledge Graph (CKG), and then learn the representations of users and items by applying graph convolutional networks to aggregate high-order neighbor information between entities in CKG. However, existing related methods mainly focus on learning representations in the Euclidean space, posing challenges in capturing the hierarchical structure and intricate relational logic between users and items. In view of this, we propose a novel hyperbolic CKG learning model HCKGL for recommendation, which leverages relation-specific curvature and attention-based geometric transformations to preserve the inherent features of CKG. Additionally, we address two significant challenges that existing methods have often overlooked. Firstly, in order to capture the relationship dependencies between neighbors and accurately calculate the contribution of neighbor information, we propose a hyperbolic graph attention network (HGAT), which combines the curvature of the relationship to assign weights. Secondly, we present a new graph contrastive learning technique (HMCL) that utilizes the hyperbolic embedding propagation and multi-level contrastive learning to improve the representations of users and items. Comprehensive experimental results on two widely used datasets demonstrate that HCKGL outperforms state-of-the-art baselines. The source code for our model is publicly available at: <span><span>https://github.com/GDM-SCNU/HCKGL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129808"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548141","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Bilinear-experts network with self-adaptive sampler for long-tailed visual recognition","authors":"Qin Wang ,&nbsp;Sam Kwong ,&nbsp;Xizhao Wang","doi":"10.1016/j.neucom.2025.129832","DOIUrl":"10.1016/j.neucom.2025.129832","url":null,"abstract":"<div><div>Long-tail distributed data hinders the practical application of state-of-the-art deep models in computer vision. Consequently, exclusive methodologies for handling the long-tailed problem are proposed, focusing on different hierarchies. For embedding hierarchy, existing works manually augment the diversity of tail-class features for specific datasets. However, prior knowledge about datasets is not always available for practical use, which brings unsatisfactory generalization ability in human fine-turned augmentation under such circumstances. To figure out this problem, we introduce a novel model named Bilinear-Experts Network (BENet) with Self-Adaptive Sampler (SAS). This model leverages model-driven perturbations to tail-class embeddings while preserving generalization capability on head classes through a designed bilinear experts system. The designed perturbations adaptively augment tail-class space and shift the class boundary away from the tail-class centers. Moreover, we find that SAS automatically assigns more significant perturbations to specific tail classes with relatively fewer training samples, which indicates SAS is capable of filtering tail classes with lower quality and enhancing them. Also, experiments conducted across various long-tailed benchmarks validate the comparable performance of the proposed BENet.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"633 ","pages":"Article 129832"},"PeriodicalIF":5.5,"publicationDate":"2025-03-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143548957","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}