Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129896
Jiaqi Luo , Yuan Yuan , Shixin Xu
Class imbalance poses a persistent challenge in machine learning, particularly for tabular data classification tasks. While Gradient Boosting Decision Trees (GBDT) models are widely regarded as state-of-the-art for these tasks, their effectiveness diminishes in the presence of imbalanced datasets. This paper is the first to comprehensively explore the integration of class-balanced loss functions into three popular GBDT algorithms, addressing binary, multi-class, and multi-label classification. We present a novel benchmark, derived from extensive experiments across diverse datasets, to evaluate the performance gains from class-balanced losses in GBDT models. Our findings establish the efficacy of these loss functions in enhancing model performance under class imbalance, providing actionable insights for practitioners tackling real-world imbalanced data challenges. To bridge the gap between research and practice, we introduce an open-source Python package that simplifies the application of class-balanced loss functions within GBDT workflows, democratizing access to these advanced methodologies. The code is available at https://github.com/Luojiaqimath/ClassbalancedLoss4GBDT.
{"title":"Improving GBDT performance on imbalanced datasets: An empirical study of class-balanced loss functions","authors":"Jiaqi Luo , Yuan Yuan , Shixin Xu","doi":"10.1016/j.neucom.2025.129896","DOIUrl":"10.1016/j.neucom.2025.129896","url":null,"abstract":"<div><div>Class imbalance poses a persistent challenge in machine learning, particularly for tabular data classification tasks. While Gradient Boosting Decision Trees (GBDT) models are widely regarded as state-of-the-art for these tasks, their effectiveness diminishes in the presence of imbalanced datasets. This paper is the first to comprehensively explore the integration of class-balanced loss functions into three popular GBDT algorithms, addressing binary, multi-class, and multi-label classification. We present a novel benchmark, derived from extensive experiments across diverse datasets, to evaluate the performance gains from class-balanced losses in GBDT models. Our findings establish the efficacy of these loss functions in enhancing model performance under class imbalance, providing actionable insights for practitioners tackling real-world imbalanced data challenges. To bridge the gap between research and practice, we introduce an open-source Python package that simplifies the application of class-balanced loss functions within GBDT workflows, democratizing access to these advanced methodologies. The code is available at <span><span>https://github.com/Luojiaqimath/ClassbalancedLoss4GBDT</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129896"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143619518","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129898
Luhuan Fei , Xinyi Wang , Jiankun Wang , Lu Sun , Yuyao Zhang
Traditional learning usually assumes that all samples share the same global model, which fails to preserve critical local information for heterogeneous data. It can be tackled by detecting sample clusters and learning sample-specific models but is limited to sample-level clustering and sample-specific feature selection. In this paper, we propose multi-level sparse network lasso (MSN Lasso) for flexible local learning. It multiplicatively decomposes model parameters into two components: One component is for coarse-grained group-level, and another is for fine-grained entry-level. At the clustering stage, MSN Lasso simultaneously groups samples (group-level) and clusters specific features across samples (entry-level). At the feature selection stage, it enables both across-sample (group-level) and sample-specific (entry-level) feature selection. Theoretical analysis reveals a potential equivalence to a jointly regularized local model, which informs the development of an efficient algorithm. A divide-and-conquer optimization strategy is further introduced to enhance the algorithm’s efficiency. Extensive experiments across diverse datasets demonstrate that MSN Lasso outperforms existing methods and exhibits greater flexibility.
{"title":"Multi-level sparse network lasso: Locally sparse learning with flexible sample clusters","authors":"Luhuan Fei , Xinyi Wang , Jiankun Wang , Lu Sun , Yuyao Zhang","doi":"10.1016/j.neucom.2025.129898","DOIUrl":"10.1016/j.neucom.2025.129898","url":null,"abstract":"<div><div>Traditional learning usually assumes that all samples share the same global model, which fails to preserve critical local information for heterogeneous data. It can be tackled by detecting sample clusters and learning sample-specific models but is limited to sample-level clustering and sample-specific feature selection. In this paper, we propose multi-level sparse network lasso (MSN Lasso) for flexible local learning. It multiplicatively decomposes model parameters into two components: One component is for coarse-grained group-level, and another is for fine-grained entry-level. At the clustering stage, MSN Lasso simultaneously groups samples (group-level) and clusters specific features across samples (entry-level). At the feature selection stage, it enables both across-sample (group-level) and sample-specific (entry-level) feature selection. Theoretical analysis reveals a potential equivalence to a jointly regularized local model, which informs the development of an efficient algorithm. A divide-and-conquer optimization strategy is further introduced to enhance the algorithm’s efficiency. Extensive experiments across diverse datasets demonstrate that MSN Lasso outperforms existing methods and exhibits greater flexibility.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"635 ","pages":"Article 129898"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143621499","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129912
Xavier Hickman, Yang Lu, Daniel Prince
Safe reinforcement learning (SRL) aims to optimize control policies that maximize long-term reward, while adhering to safety constraints. SRL has many real-world applications such as, autonomous vehicles, industrial robotics, and healthcare. Recent advances in offline reinforcement learning (RL) — where agents learn policies from static datasets without interacting with the environment — have made it a promising approach to derive safe control policies. However, offline RL faces significant challenges, such as covariate shift and outliers in the data, which can lead to suboptimal policies. Similarly, online SRL, which derives safe policies through real-time environment interaction, struggles with outliers and often relies on unrealistic regularity assumptions, limiting its practicality. This paper addresses these challenges by proposing a hybrid-offline–online approach. First, prior knowledge from offline learning guides online exploration. Then, during online learning, we replace the popular Gaussian Process (GP) with the Student-t’s Process (TP) to enhance robustness to covariate shift and outliers.
{"title":"Hybrid safe reinforcement learning: Tackling distribution shift and outliers with the Student-t’s process","authors":"Xavier Hickman, Yang Lu, Daniel Prince","doi":"10.1016/j.neucom.2025.129912","DOIUrl":"10.1016/j.neucom.2025.129912","url":null,"abstract":"<div><div>Safe reinforcement learning (SRL) aims to optimize control policies that maximize long-term reward, while adhering to safety constraints. SRL has many real-world applications such as, autonomous vehicles, industrial robotics, and healthcare. Recent advances in offline reinforcement learning (RL) — where agents learn policies from static datasets without interacting with the environment — have made it a promising approach to derive safe control policies. However, offline RL faces significant challenges, such as covariate shift and outliers in the data, which can lead to suboptimal policies. Similarly, online SRL, which derives safe policies through real-time environment interaction, struggles with outliers and often relies on unrealistic regularity assumptions, limiting its practicality. This paper addresses these challenges by proposing a hybrid-offline–online approach. First, prior knowledge from offline learning guides online exploration. Then, during online learning, we replace the popular Gaussian Process (GP) with the Student-t’s Process (TP) to enhance robustness to covariate shift and outliers.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129912"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143610917","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129884
Mingsen Du , Yanxuan Wei , Xiangwei Zheng , Cun Ji
Recently, time series classification has attracted the attention of a large number of researchers, and hundreds of methods have been proposed. However, these methods often ignore the spatial correlations among dimensions and the local correlations among features. To address this issue, the causal and local correlations based network (CaLoNet) is proposed in this study for multivariate time series classification. First, pairwise spatial correlations between dimensions are modeled using causality modeling to obtain the graph structure. Then, a relationship extraction network is used to fuse local correlations to obtain long-term dependency features. Finally, the graph structure and long-term dependency features are integrated into the graph neural network. Experiments on the UEA datasets show that CaLoNet can obtain competitive performance compared with state-of-the-art methods.
{"title":"Causal and Local Correlations Based Network for Multivariate Time Series Classification","authors":"Mingsen Du , Yanxuan Wei , Xiangwei Zheng , Cun Ji","doi":"10.1016/j.neucom.2025.129884","DOIUrl":"10.1016/j.neucom.2025.129884","url":null,"abstract":"<div><div>Recently, time series classification has attracted the attention of a large number of researchers, and hundreds of methods have been proposed. However, these methods often ignore the spatial correlations among dimensions and the local correlations among features. To address this issue, the causal and local correlations based network (CaLoNet) is proposed in this study for multivariate time series classification. First, pairwise spatial correlations between dimensions are modeled using causality modeling to obtain the graph structure. Then, a relationship extraction network is used to fuse local correlations to obtain long-term dependency features. Finally, the graph structure and long-term dependency features are integrated into the graph neural network. Experiments on the UEA datasets show that CaLoNet can obtain competitive performance compared with state-of-the-art methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129884"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143619556","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129899
Robert Guamán-Rivera , Jose Delpiano , Rodrigo Verschae
Developing new convolutional neural network architectures and event-based camera representations could play a crucial role in autonomous navigation, pose estimation, and visual odometry applications. This study explores the potential of event cameras in optical flow estimation using convolutional neural networks. We provide a detailed description of the principles of operation and the software available for extracting and processing information from event cameras, along with the various event representation methods offered by this technology. Likewise, we identify four method categories to estimate optical flow using event cameras: gradient-based, frequency-based, correlation-based and neural network models. We report on these categories, including their latest developments, current status and challenges. We provide information on existing datasets and identify the appropriate dataset to evaluate deep learning-based optical flow estimation methods. We evaluate the accuracy of the implemented methods using the average endpoint error metric; meanwhile, the efficiency of the algorithms is evaluated as a function of execution time. Finally, we discuss research directions that promise future advances in this field.
{"title":"Event-based optical flow: Method categorisation and review of techniques that leverage deep learning","authors":"Robert Guamán-Rivera , Jose Delpiano , Rodrigo Verschae","doi":"10.1016/j.neucom.2025.129899","DOIUrl":"10.1016/j.neucom.2025.129899","url":null,"abstract":"<div><div>Developing new convolutional neural network architectures and event-based camera representations could play a crucial role in autonomous navigation, pose estimation, and visual odometry applications. This study explores the potential of event cameras in optical flow estimation using convolutional neural networks. We provide a detailed description of the principles of operation and the software available for extracting and processing information from event cameras, along with the various event representation methods offered by this technology. Likewise, we identify four method categories to estimate optical flow using event cameras: gradient-based, frequency-based, correlation-based and neural network models. We report on these categories, including their latest developments, current status and challenges. We provide information on existing datasets and identify the appropriate dataset to evaluate deep learning-based optical flow estimation methods. We evaluate the accuracy of the implemented methods using the average endpoint error metric; meanwhile, the efficiency of the algorithms is evaluated as a function of execution time. Finally, we discuss research directions that promise future advances in this field.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"635 ","pages":"Article 129899"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143628890","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129834
Haowen Xu , Mingwei Tang , Tao Cai , Jie Hu , Zhongyuan Jiang , Deng Bian , Shixuan Lv
Aspect Sentiment Triplet Extraction (ASTE) is a method for extracting aspect terms, opinion terms, and their corresponding sentiment polarities from a given sentence. Most of the existing studies use joint extraction methods to extract the triplets directly in a unified framework. However, most joint extraction methods only consider the semantic and syntactic dependency information of the sentence. Due to a lack of sentiment information and positional information, they are unable to accurately and completely express the aspect and opinion in the sentence. In order to solve the above problems, we introduce a Multiple-level Enhanced Graph Convolutional Network (MEGCN) for ASTE, which utilizes sentiment scores and sentiment polarity nodes alongside syntactic dependency information. This approach not only enriches contextual understanding by integrating sentiment data but also improves positional analysis of aspect and opinion terms through polarity nodes. Moreover, our dual-aware fusion module, combining semantic with sentiment-enhanced syntactic features through a biaffine attention mechanism and matrix construction, enables a deeper representation of aspect sentiment triplets. Our model demonstrates superior performance over existing methods on two widely recognized ASTE datasets.
{"title":"Multiple-level Enhanced Graph Convolutional Network for Aspect Sentiment Triplet Extraction","authors":"Haowen Xu , Mingwei Tang , Tao Cai , Jie Hu , Zhongyuan Jiang , Deng Bian , Shixuan Lv","doi":"10.1016/j.neucom.2025.129834","DOIUrl":"10.1016/j.neucom.2025.129834","url":null,"abstract":"<div><div>Aspect Sentiment Triplet Extraction (ASTE) is a method for extracting aspect terms, opinion terms, and their corresponding sentiment polarities from a given sentence. Most of the existing studies use joint extraction methods to extract the triplets directly in a unified framework. However, most joint extraction methods only consider the semantic and syntactic dependency information of the sentence. Due to a lack of sentiment information and positional information, they are unable to accurately and completely express the aspect and opinion in the sentence. In order to solve the above problems, we introduce a Multiple-level Enhanced Graph Convolutional Network (MEGCN) for ASTE, which utilizes sentiment scores and sentiment polarity nodes alongside syntactic dependency information. This approach not only enriches contextual understanding by integrating sentiment data but also improves positional analysis of aspect and opinion terms through polarity nodes. Moreover, our dual-aware fusion module, combining semantic with sentiment-enhanced syntactic features through a biaffine attention mechanism and matrix construction, enables a deeper representation of aspect sentiment triplets. Our model demonstrates superior performance over existing methods on two widely recognized ASTE datasets.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129834"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143619519","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129918
Jaesin Ahn, Heechul Jung
Deep learning-based MRI reconstruction methods have gained significant attention recently due to the need for accelerated MRI scans. However, existing deep learning-based methods for off-resonance correction rely on simple CNNs, resulting in suboptimal solutions. In this paper, we propose a gated dual domain transformer with gated spatial projection and gated frequency projection to effectively handle complex-valued MRI, as the first attempt to utilize transformer-based model for off-resonance correction. Additionally, we introduce a selective perceptual loss with a novel test-time translation-merger to reconstruct perceptually high-quality images without checkerboard artifacts. Experiments on both simulated and real off-resonance MRI datasets demonstrate the effectiveness of our approach. Furthermore, we also present ablation studies to determine the optimal design choices.
{"title":"GDoT: A gated dual domain transformer for enhanced MRI off-resonance correction","authors":"Jaesin Ahn, Heechul Jung","doi":"10.1016/j.neucom.2025.129918","DOIUrl":"10.1016/j.neucom.2025.129918","url":null,"abstract":"<div><div>Deep learning-based MRI reconstruction methods have gained significant attention recently due to the need for accelerated MRI scans. However, existing deep learning-based methods for off-resonance correction rely on simple CNNs, resulting in suboptimal solutions. In this paper, we propose a gated dual domain transformer with gated spatial projection and gated frequency projection to effectively handle complex-valued MRI, as the first attempt to utilize transformer-based model for off-resonance correction. Additionally, we introduce a selective perceptual loss with a novel test-time translation-merger to reconstruct perceptually high-quality images without checkerboard artifacts. Experiments on both simulated and real off-resonance MRI datasets demonstrate the effectiveness of our approach. Furthermore, we also present ablation studies to determine the optimal design choices.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129918"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143610915","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}
Pub Date : 2025-03-11DOI: 10.1016/j.neucom.2025.129916
Aakash Kumar , Lei Zhang , Hazrat Bilal , Shifeng Wang , Ali Muhammad Shaikh , Lu Bo , Avinash Rohra , Alisha Khalid
With the rapid advancement of science and technology, the field of mobile robot applications continues to expand, with path planning emerging as a fundamental yet challenging task. While traditional path planning techniques have developed into a relatively complete theoretical system, their performance in uncertain environments remains a critical area of research. To address this, we propose a novel Deep Spiking Q-Network (DSQN) algorithm that significantly enhances path planning performance by leveraging the unique advantages of spiking neural networks (SNNs). Unlike classic Q-learning and its contemporary variants, the DSQN algorithm integrates global and local information simultaneously, resulting in superior overall performance. As the third generation of neural networks, SNNs offer unparalleled robustness and energy efficiency by mimicking biological neural systems. By introducing spiking neurons into the conventional Deep Q-learning (DQN) framework, the DSQN algorithm overcomes key challenges in deep reinforcement learning (DRL), such as limited robustness and high energy consumption. The DSQN training process incorporates both surrogate gradient learning (SGL) and ANN-to-SNN conversion techniques, with SGL demonstrating remarkable effectiveness in mobile robot path planning tasks. Experimental results validate the practicality and efficiency of DSQN, showcasing improved performance across diverse test scenarios compared to the original DQN algorithm. These findings highlight the potential of DSQN to advance path planning in complex and uncertain environments, establishing it as a robust and energy-efficient solution for mobile robotics.
{"title":"DSQN: Robust path planning of mobile robot based on deep spiking Q-network","authors":"Aakash Kumar , Lei Zhang , Hazrat Bilal , Shifeng Wang , Ali Muhammad Shaikh , Lu Bo , Avinash Rohra , Alisha Khalid","doi":"10.1016/j.neucom.2025.129916","DOIUrl":"10.1016/j.neucom.2025.129916","url":null,"abstract":"<div><div>With the rapid advancement of science and technology, the field of mobile robot applications continues to expand, with path planning emerging as a fundamental yet challenging task. While traditional path planning techniques have developed into a relatively complete theoretical system, their performance in uncertain environments remains a critical area of research. To address this, we propose a novel Deep Spiking Q-Network (DSQN) algorithm that significantly enhances path planning performance by leveraging the unique advantages of spiking neural networks (SNNs). Unlike classic Q-learning and its contemporary variants, the DSQN algorithm integrates global and local information simultaneously, resulting in superior overall performance. As the third generation of neural networks, SNNs offer unparalleled robustness and energy efficiency by mimicking biological neural systems. By introducing spiking neurons into the conventional Deep Q-learning (DQN) framework, the DSQN algorithm overcomes key challenges in deep reinforcement learning (DRL), such as limited robustness and high energy consumption. The DSQN training process incorporates both surrogate gradient learning (SGL) and ANN-to-SNN conversion techniques, with SGL demonstrating remarkable effectiveness in mobile robot path planning tasks. Experimental results validate the practicality and efficiency of DSQN, showcasing improved performance across diverse test scenarios compared to the original DQN algorithm. These findings highlight the potential of DSQN to advance path planning in complex and uncertain environments, establishing it as a robust and energy-efficient solution for mobile robotics.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129916"},"PeriodicalIF":5.5,"publicationDate":"2025-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143619557","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}
Pub Date : 2025-03-10DOI: 10.1016/j.neucom.2025.129892
John Atanbori , Christos A. Frantzidis , Mohammed Al-Khafajiy , Aliyu Aliyu , Behnaz Sohani , Kofi Appiah , Harriet Moore , Catherine Sanders , Alastair I. Ward
The identification of biological echoes in radar data has revolutionized research into airborne migratory species. Deep learning applied to polarimetric weather radar observations can reveal signature patterns of mass movement by bio-scatterers such as birds, bats, and insects. However, due to the difficulties in labelling bio-scatterers in these data, threshold approaches have been proposed in the literature. In this research, we used the depolarization ratio (DR) based on differential reflectivity (zDR) and the cross-correlation coefficient (pHV), along with citizen scientist-reported data, to label bio-scatterers for deep learning. This method of labelling biological echoes in radar signatures is prone to noise, which impacts the accuracy of any model that relies on it. We introduce a novel semi-supervised co-training approach that uses a bootstrap ensemble with a confidence threshold. Our ensemble consists of the newly proposed STNet and two modified FNet models, which incorporate co-learning through bootstrap sampling for label correction. This innovative method significantly improves classification accuracy across all three multivariate numerical datasets compared to baseline models that lack co-learning with bootstrap-based label correction.
{"title":"Learning with noisy labels for classifying biological echoes in polarimetric weather radar observations using artificial neural networks","authors":"John Atanbori , Christos A. Frantzidis , Mohammed Al-Khafajiy , Aliyu Aliyu , Behnaz Sohani , Kofi Appiah , Harriet Moore , Catherine Sanders , Alastair I. Ward","doi":"10.1016/j.neucom.2025.129892","DOIUrl":"10.1016/j.neucom.2025.129892","url":null,"abstract":"<div><div>The identification of biological echoes in radar data has revolutionized research into airborne migratory species. Deep learning applied to polarimetric weather radar observations can reveal signature patterns of mass movement by bio-scatterers such as birds, bats, and insects. However, due to the difficulties in labelling bio-scatterers in these data, threshold approaches have been proposed in the literature. In this research, we used the depolarization ratio (DR) based on differential reflectivity (zDR) and the cross-correlation coefficient (pHV), along with citizen scientist-reported data, to label bio-scatterers for deep learning. This method of labelling biological echoes in radar signatures is prone to noise, which impacts the accuracy of any model that relies on it. We introduce a novel semi-supervised co-training approach that uses a bootstrap ensemble with a confidence threshold. Our ensemble consists of the newly proposed STNet and two modified FNet models, which incorporate co-learning through bootstrap sampling for label correction. This innovative method significantly improves classification accuracy across all three multivariate numerical datasets compared to baseline models that lack co-learning with bootstrap-based label correction.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"634 ","pages":"Article 129892"},"PeriodicalIF":5.5,"publicationDate":"2025-03-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143610916","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}
Pub Date : 2025-03-10DOI: 10.1016/j.neucom.2025.129633
Han Ren , Xiaona Chang , Xia Li
Automatic headline generation (HG) is an important natural language processing (NLP) task that aims to obtain a highly compressed text snippet from a document, to exhibit the core concept. Traditional headline generation (HG) techniques predominantly employ text summarization methods to generate short texts, by selecting important information from original documents. In recent years, with the rapid development of deep learning techniques, research on HG has leaned toward neural network-based end-to-end modeling approaches. Pretrained schemes and large language models (LLMs) demonstrate superior capability in generating natural language texts, thereby promoting further exploration on HG studies. However, a quality gap remains between machine-generated and human-written texts, making the generation of attractive and faithful headlines worthy of in-depth research. Therefore, this study presents a review of the most recent technologies on HG, including methods, datasets, and evaluation strategies. Future research directions are outlined, which provide a valuable reference point for HG and other summarization tasks. A collection of reference papers and code sources is available at: https://github.com/xiaona-chang/HGSurvey.
{"title":"Neural headline generation: A comprehensive survey","authors":"Han Ren , Xiaona Chang , Xia Li","doi":"10.1016/j.neucom.2025.129633","DOIUrl":"10.1016/j.neucom.2025.129633","url":null,"abstract":"<div><div>Automatic headline generation (HG) is an important natural language processing (NLP) task that aims to obtain a highly compressed text snippet from a document, to exhibit the core concept. Traditional headline generation (HG) techniques predominantly employ text summarization methods to generate short texts, by selecting important information from original documents. In recent years, with the rapid development of deep learning techniques, research on HG has leaned toward neural network-based end-to-end modeling approaches. Pretrained schemes and large language models (LLMs) demonstrate superior capability in generating natural language texts, thereby promoting further exploration on HG studies. However, a quality gap remains between machine-generated and human-written texts, making the generation of attractive and faithful headlines worthy of in-depth research. Therefore, this study presents a review of the most recent technologies on HG, including methods, datasets, and evaluation strategies. Future research directions are outlined, which provide a valuable reference point for HG and other summarization tasks. A collection of reference papers and code sources is available at: <span><span>https://github.com/xiaona-chang/HGSurvey</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"635 ","pages":"Article 129633"},"PeriodicalIF":5.5,"publicationDate":"2025-03-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143628886","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号