Neurocomputing最新文献

{"title":"MCI-GRU: Stock prediction model based on multi-head cross-attention and improved GRU","authors":"Peng Zhu ,&nbsp;Yuante Li ,&nbsp;Yifan Hu ,&nbsp;Sheng Xiang ,&nbsp;Qinyuan Liu ,&nbsp;Dawei Cheng ,&nbsp;Yuqi Liang","doi":"10.1016/j.neucom.2025.130168","DOIUrl":"10.1016/j.neucom.2025.130168","url":null,"abstract":"<div><div>As financial markets become increasingly complex and the era of big data unfolds, accurate stock prediction has become more critical. Although traditional time series models, such as GRU, have been widely applied to stock prediction, they still exhibit limitations in addressing the intricate nonlinear dynamics of markets, particularly in the flexible selection and effective utilization of key historical information. In recent years, emerging methods like Graph Neural Networks and Reinforcement Learning have shown significant potential in stock prediction. However, these methods often demand high data quality and quantity, and they tend to exhibit instability when dealing with data sparsity and noise. Moreover, the training and inference processes for these models are typically complex and computationally expensive, limiting their broad deployment in practical applications. Existing approaches also generally struggle to capture unobservable latent market states effectively, such as market sentiment and expectations, microstructural factors, and participant behavior patterns, leading to an inadequate understanding of market dynamics and subsequently impact prediction accuracy. To address these challenges, this paper proposes a stock prediction model, MCI-GRU, based on a multi-head cross-attention mechanism and an improved GRU. First, we enhance the GRU model by replacing the reset gate with an attention mechanism, thereby increasing the model’s flexibility in selecting and utilizing historical information. Second, we design a multi-head cross-attention mechanism for learning unobservable latent market state representations, which are further enriched through interactions with both temporal features and cross-sectional features. Finally, extensive experiments conducted on the CSI 300 and CSI 500 datasets from the Chinese stock market, as well as the NASDAQ 100 and S&amp;P 500 datasets from the U.S. stock market, demonstrate that the proposed method outperforms the current state-of-the-art methods across multiple metrics. Furthermore, this approach has been successfully applied in the real-world operations of a fund management company, validating its effectiveness and practicality in actual financial environments.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130168"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143785315","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":"Debate divides: Argument relation-based contrastive opinion summarization via multi-task learning for online discussions","authors":"Huawei Shan ,&nbsp;Dongyuan Lu","doi":"10.1016/j.neucom.2025.130124","DOIUrl":"10.1016/j.neucom.2025.130124","url":null,"abstract":"<div><div>Contrastive opinion summarization(COS) aims to generate summaries that can see both sides of an argument. Existing studies mostly focus on summarizing popular opinions expressed in comments, often ignoring divergent opinions. While some studies use sentiment-based frameworks to compare opinions on the same topic, relying solely on sentiment polarity proves inadequate for distinguishing nuanced differences in opinion. Additionally, many methods extract topics directly from comments without considering their relevance to the associated reading content. In this paper, we propose a novel Topic-level Contrastive Opinion Summarization model for online discussions, called ARCOS, which leverages argument relations between comments to identify divergent opinion pairs. Inspired by cognitive map theory, ARCOS first extracts topics from a keyword co-occurrence graph generated from the reading content. It then uses a multi-task learning network to predict argument relations between comments and align them with the extracted content topics. Based on argument relations, ARCOS selects representative contrastive comment pairs from comments on the same topic, and employs a large language model to produce a divergent opinion summary. Experimental results on a newly created benchmark dataset, RNCOS, show that ARCOS outperforms baselines in various sub-tasks and generates high-quality summaries for divergent opinions about specific content topics.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130124"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143791493","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":"Stage-wise multi-focus fusion for numerous misaligned industrial images","authors":"Wenjie Zhu ,&nbsp;Xurong Chi ,&nbsp;Jingrun Chen","doi":"10.1016/j.neucom.2025.130062","DOIUrl":"10.1016/j.neucom.2025.130062","url":null,"abstract":"<div><div>Multi-focus image fusion is a technique that combines information from multiple images to generate a single composite image that retains all the essential details and features of the original images. Traditional methods can achieve rapid fusion but struggle with image misalignment. Due to the strong expressive power of deep neural networks, deep learning methods can handle misaligned situations but face challenges when fusing many images. In this work, a stage-wise multi-focus fusion method is proposed. Firstly, in the coarse fusion stage, the number of images for subsequent fusion is reduced by fusing similar images. Here, several consecutive images are sequentially fused by using an indicator to distinguish between clear and blurred areas and then merging the clear regions. Subsequently, in the fine fusion stage, pairwise merging is adopted to obtain the final result. For each pair of sub-images, a blur filter, difference operator, and guided filter are utilized to create a decision map, followed by pixel-wise weighted averaging to fuse the source images. The proposed method is validated using two multi-focus image datasets with numerous misaligned industrial images, namely the Lampwick dataset and the Impurity dataset. Moreover, its generalization ability is demonstrated by conducting experiments on the Lytro and MFFW datasets. Our findings show that, despite the industrial images often possessing the characteristics of being numerous and misaligned, our method not only achieves fast fusion but also yields the best results.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130062"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143800078","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":"PT-VAE: Variational autoencoder with prior concept transformation","authors":"Zitu Liu ,&nbsp;Yue Liu ,&nbsp;Zhenyao Yu ,&nbsp;Zhengwei Yang ,&nbsp;Qingshan Fu ,&nbsp;Yike Guo ,&nbsp;Qun Liu ,&nbsp;Guoyin Wang","doi":"10.1016/j.neucom.2025.130129","DOIUrl":"10.1016/j.neucom.2025.130129","url":null,"abstract":"<div><div>Learning and disentangling coherent latent representations of variational autoencoders (VAEs) have recently attracted widespread attention. However, the latent space of the VAE model is constrained by the prior distribution, which can hinder the latent variables from accurately capturing semantic information, thereby limiting its disentanglement and interpretability. This paper proposes PT-VAE, which constructs the latent space by a well-constructed latent space rather than a carefully designed prior distribution to guide the latent variables. Firstly, we transform the initial constraints of the latent space into understandable latent variable distributions, the so-called prior concept, which can be introduced into the latent space. Then, we design the Gumbel softmax reparameterization trick to enhance the integration of the prior concept and latent variables. Furthermore, the training process of PT-VAE is guided by deriving a variational lower bound, which facilitates the construction of the latent space concept based on the prior concept. Compared with 8 state-of-the-art VAE models, the PT-VAE improves the average clustering accuracy by over 11 % on the Fashion MNSIT, MNIST, COIL20, and COIL10 datasets. Moreover, the PT-VAE elucidates the process of information aggregation within the model and uncovers disentangled representations. PT-VAE provides a novel and flexible approach to construct an interpretable latent space by embedding prior concepts and disentangling the latent variables.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130129"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143817512","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 reinforcement learning approach for optimized MRI sampling with region-specific fidelity","authors":"Ruru Xu,&nbsp;Ilkay Oksuz","doi":"10.1016/j.neucom.2025.130116","DOIUrl":"10.1016/j.neucom.2025.130116","url":null,"abstract":"<div><div>Accelerating Magnetic Resonance Imaging (MRI) acquisition while preserving diagnostic quality remains a significant challenge in medical imaging. This paper introduces a novel reinforcement learning approach for optimizing k-space sampling, addressing the critical need for maintaining high-fidelity reconstructions, particularly in clinically significant regions. Our method bridges k-space and image domains by employing a multi-layer Fast Fourier Transform (FFT) network architecture coupled with a comprehensive reward function. The reward function integrates global SSIM, region-specific metrics, and k-space MSE, achieving SSIM scores of 0.9764 (global) and PSNR of 43.427 in cardiac imaging, with a Dice score of 0.9606 for critical regions. This function uniquely balances global image fidelity with region-specific accuracy, ensuring optimal sampling in areas of high diagnostic value. We present a reinforcement learning framework that maintains consistency with MRI physics throughout the optimization process by utilizing Proximal Policy Optimization (PPO) for concurrent refinement of policy and value networks, while minimizing unnecessary domain transformations. We validate our approach on two diverse datasets: the ACDC cardiac MRI and the FastMRI knee dataset, demonstrating its effectiveness across different regions of clinical interest and showing consistent improvements across 4x and 6x acceleration factors. The proposed k-space sampling optimization technique not only enhances reconstruction quality but also serves as a versatile front-end tool, adaptable to various MRI reconstruction algorithms. The code is publicly available at <span><span>https://github.com/Ruru-Xu/KSRO</span><svg><path></path></svg></span></div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130116"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143817506","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":"TransCMFD: An adaptive transformer for copy-move forgery detection","authors":"Enji Liang ,&nbsp;Kuiyuan Zhang ,&nbsp;Zhongyun Hua ,&nbsp;Yuanman Li ,&nbsp;Xiaohua Jia","doi":"10.1016/j.neucom.2025.130110","DOIUrl":"10.1016/j.neucom.2025.130110","url":null,"abstract":"<div><div>Copy-move forgery is one of the most common image tampering methods. It is a frequently employed method for manipulating evidence or deceiving the public by hiding some objects in an image or replicating significant objects. Therefore, it is crucial to focus on copy-move forgery detection. In this paper, we propose TransCMFD as a new transformer-based model for detecting copy-move forgery. We propose an adaptive transformer encoder and combine the traditional convolution encoder–decoder to capture different global and local features of the forgery image, respectively. This can enhance the model’s comprehension of the impact of forgery across different tampered image regions. To allow the model to concentrate more on the tampered regions that resemble the original regions, we introduce a similarity detection module. Moreover, to enhance the localization accuracy of the tampered regions, we design an adaptive loss function combination strategy that incorporates the DICE coefficient loss and binary cross-entropy loss. We perform comprehensive experiments on both synthetic and four publicly available datasets. The results show that our model has better performance in copy-move forgery detection compared to baseline methods, and it has remarkable robustness to some common image attacks such as noise addition attacks, image blurring attacks, and color reduction attacks.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130110"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143825923","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":"Be your own doctor: Temperature scaling self-knowledge distillation for medical image classification","authors":"Wenjie Liu,&nbsp;Lei Zhang,&nbsp;Xianliang Zhang,&nbsp;Xinyang Zhou,&nbsp;Xin Wei","doi":"10.1016/j.neucom.2025.130115","DOIUrl":"10.1016/j.neucom.2025.130115","url":null,"abstract":"<div><div>Self-knowledge distillation (self-KD), which uses the student network as the teacher model, allows the model to learn knowledge by itself. It has been widely studied in various medical image tasks for constructing lightweight models to alleviate the limitations of computing resources. However, existing self-KD methods use a single temperature for distillation, ignoring the effect of temperature on different classes. In this paper, we investigate the effects of target class temperature and non-target class temperature on the performance of self-KD. Based on the above study, a temperature scaling self-knowledge distillation (TSS-KD) model is proposed, which can better balance the target class knowledge and non-target class knowledge. By adjusting the temperature scaling of different classes, the model can learn better representations by distilling the well-proportioned features. To make the network focus more on the local lesions of medical images, a regional gamma augmentation (RGA) method is proposed, which provides stronger perturbations to the same sample to generate more differentiated features. By self-regularizing the consistency of these features, the model can learn more local knowledge. To evaluate the effectiveness of the proposed method, extensive experiments are conducted on nine medical image classification tasks of eight public datasets. Experimental results show that the proposed method outperforms state-of-the-art self-KD models and has strong generality. The code is available at <span><span>https://github.com/JeaneyLau/TSS-KD</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130115"},"PeriodicalIF":5.5,"publicationDate":"2025-04-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143791531","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":"Can large language models independently complete tasks? A dynamic evaluation framework for multi-turn task planning and completion","authors":"Jun Gao ,&nbsp;Junlin Cui ,&nbsp;Huijia Wu ,&nbsp;Liuyu Xiang ,&nbsp;Han Zhao ,&nbsp;Xiangang Li ,&nbsp;Meng Fang ,&nbsp;Yaodong Yang ,&nbsp;Zhaofeng He","doi":"10.1016/j.neucom.2025.130135","DOIUrl":"10.1016/j.neucom.2025.130135","url":null,"abstract":"<div><div>Large language models (LLMs) are increasingly relied upon for multi-turn dialogue to conduct complex tasks. However, existing benchmarks mainly evaluate LLMs as agents, overlooking their potential as independent systems to accomplish complex tasks. In addition, these benchmarks typically evaluate the planning and completion capabilities of the models individually, rather than simultaneously. To address these issues, we propose a new <strong>Dynamic Evaluation Framework for Multi-Turn task planning and completion (DEF-MT)</strong> to assess the ability of LLM to independently complete complex tasks in multi-turn scenarios. Our approach quantifies the model’s planning capability by guiding it to generate planning and responses sequentially. Simultaneously, we use a dynamic approach to generate data that simulates the complex intents of real users. Finally, experiments conducted on 9 mainstream models using the Multiwoz 2.2 dataset, indicate that the existing models’ sub-task planning capabilities hinder their ability to complete complex tasks, providing a meaningful reference for the future optimization direction of LLM.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130135"},"PeriodicalIF":5.5,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143783889","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":"FourierLoss: Shape-aware loss function with Fourier descriptors","authors":"Mehmet Bahadir Erden ,&nbsp;Selahattin Cansiz ,&nbsp;Onur Caki ,&nbsp;Haya Khattak ,&nbsp;Durmus Etiz ,&nbsp;Melek Cosar Yakar ,&nbsp;Kerem Duruer ,&nbsp;Berke Barut ,&nbsp;Cigdem Gunduz-Demir","doi":"10.1016/j.neucom.2025.130155","DOIUrl":"10.1016/j.neucom.2025.130155","url":null,"abstract":"<div><div>Encoder–decoder networks are commonly used for medical image segmentation tasks. When they are trained with a standard loss function, these networks are not explicitly enforced to preserve the shape integrity of an object in an image. However, this ability of the network is important to obtain more accurate results, especially when there is a low-contrast difference between the object and its surroundings. To respond this issue, this work introduces a new shape-aware loss function, which we name <em>FourierLoss</em>. This loss function relies on quantifying the shape dissimilarity between the ground truth and the predicted segmentation maps through the Fourier descriptors calculated on the objects of these maps, and penalizing this dissimilarity in network training. Different than the previous studies, <em>FourierLoss</em> offers an adaptive loss function with trainable hyperparameters that control the importance of the level of the shape details that the network is enforced to learn in the training process. This control is achieved by the proposed adaptive loss update mechanism, which end-to-end learns the hyperparameters simultaneously with the network weights by backpropagation. As a result of using this mechanism, the network can dynamically change its attention from learning the general outline of an object to learning the details of its contour points, or vice versa, in different training epochs. Working on two different datasets, our experiments revealed that the proposed adaptive shape-aware loss function led to statistically significantly better results for liver segmentation, compared to its counterparts.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130155"},"PeriodicalIF":5.5,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143800190","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":"Learning deep feature representations for multi-modal MR brain tumor segmentation","authors":"Tongxue Zhou ,&nbsp;Zheng Wang ,&nbsp;Xiaohui Liu ,&nbsp;Weibo Liu ,&nbsp;Shan Zhu","doi":"10.1016/j.neucom.2025.130162","DOIUrl":"10.1016/j.neucom.2025.130162","url":null,"abstract":"<div><div>Brain tumor segmentation is crucial for accurate diagnosis, treatment planning, and patient monitoring. Different MRI sequences can provide unique and complementary information about various aspects of brain tumors. However, effectively integrating diverse data sources to achieve accurate segmentation remains a significant challenge due to the inherent complexity and variability of the data. To address this challenge, this paper proposes a deep learning framework designed to fuse multi-modal MRI data and enhance brain tumor segmentation accuracy. Specifically, the framework introduces two innovative modules: the modality-wise feature fusion module (MFFM) and the spatial and channel-wise feature fusion module (SCFFM). The MFFM aims to learn modality-specific features and integrate information from diverse modalities, thereby ensuring richer and more discriminative feature representations. Meanwhile, the SCFFM is designed to capture contextual information and achieve multi-channel data incorporation by emphasizing informative regions and highlighting critical features. Together, these modules collaboratively enhance the model’s capacity for feature learning, leading to more precise tumor segmentation. Experimental validation on two public datasets demonstrates the effectiveness of the proposed approach, achieving an average Dice similarity coefficient of 83.2% with an average 95% Hausdorff distance of 4.3 mm on the BraTS 2018 dataset, and an average Dice similarity coefficient of 82.9% with an average 95% Hausdorff distance of 5.5 mm on the BraTS 2019 dataset. This framework not only presents an effective method for precise multi-modal brain tumor segmentation but also provides a promising solution for other challenges in multi-modal data fusion.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"638 ","pages":"Article 130162"},"PeriodicalIF":5.5,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143821337","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}