Machine learning with applications最新文献_第8页

Model-informed oracle training for enhancing active learning without external knowledge 基于模型的oracle训练，在没有外部知识的情况下增强主动学习

IF 4.9

Machine learning with applications

Pub Date : 2025-10-27 DOI: 10.1016/j.mlwa.2025.100775

Yujin Cha

In real-world applications of active learning frameworks, human oracles are often imperfect, and label noise is introduced into the learning process. This issue can be mitigated by further training the oracle using previous knowledge acquired by the model. However, it remains unclear whether model-informed oracle training can significantly improve performance. This study investigates whether recursive feedback between the model and the oracle can induce a knowledge augmentation effect, defined as a statistically significant improvement in model performance after receiving feedback from a self-data-trained oracle. To this end, we implemented a bidirectional active learning framework in which the model assists oracle learning by selectively transferring prior knowledge. In a closed-loop environment without external data, the model performs informative sample selection from an unlabeled pool, querying the oracle for labels, and retraining on the updated dataset. Simultaneously, the oracle is updated by learning from samples from the model’s training data that exhibit high uncertainty from the oracle’s perspective. This framework was empirically validated through a behavioral experiment involving 252 clinicians performing a medical image interpretation task. The results showed that model-informed oracle training enhanced both oracle accuracy and model performance. Moreover, when oracle learning was constrained by a fixed learning budget, a sampling strategy jointly balancing uncertainty and representativeness yielded the strongest effect. These findings provide compelling empirical evidence of the knowledge augmentation effect arising from human learning within a closed-loop active learning framework.

在主动学习框架的实际应用中，人类的预言器通常是不完美的，并且在学习过程中引入了标签噪声。这个问题可以通过使用模型获得的先前知识进一步训练oracle来缓解。然而，目前还不清楚基于模型的oracle训练是否能显著提高性能。本研究探讨了模型和oracle之间的递归反馈是否可以诱导知识增强效应，即在接受自数据训练的oracle的反馈后，模型性能在统计上显着提高。为此，我们实现了一个双向主动学习框架，其中模型通过选择性地转移先验知识来帮助oracle学习。在没有外部数据的闭环环境中，模型从未标记的池中执行信息样本选择，向oracle查询标签，并对更新的数据集进行再训练。同时，通过从模型的训练数据中学习样本来更新oracle，从oracle的角度来看，这些数据显示出很高的不确定性。通过252名临床医生执行医学图像解释任务的行为实验，该框架得到了经验验证。结果表明，基于模型的oracle训练提高了oracle的准确率和模型性能。此外，当oracle学习受到固定学习预算的约束时，联合平衡不确定性和代表性的抽样策略效果最强。这些发现提供了令人信服的经验证据，证明在闭环主动学习框架下，人类学习产生的知识增强效应。

{"title":"Model-informed oracle training for enhancing active learning without external knowledge","authors":"Yujin Cha","doi":"10.1016/j.mlwa.2025.100775","DOIUrl":"10.1016/j.mlwa.2025.100775","url":null,"abstract":"<div><div>In real-world applications of active learning frameworks, human oracles are often imperfect, and label noise is introduced into the learning process. This issue can be mitigated by further training the oracle using previous knowledge acquired by the model. However, it remains unclear whether model-informed oracle training can significantly improve performance. This study investigates whether recursive feedback between the model and the oracle can induce a knowledge augmentation effect, defined as a statistically significant improvement in model performance after receiving feedback from a self-data-trained oracle. To this end, we implemented a bidirectional active learning framework in which the model assists oracle learning by selectively transferring prior knowledge. In a closed-loop environment without external data, the model performs informative sample selection from an unlabeled pool, querying the oracle for labels, and retraining on the updated dataset. Simultaneously, the oracle is updated by learning from samples from the model’s training data that exhibit high uncertainty from the oracle’s perspective. This framework was empirically validated through a behavioral experiment involving 252 clinicians performing a medical image interpretation task. The results showed that model-informed oracle training enhanced both oracle accuracy and model performance. Moreover, when oracle learning was constrained by a fixed learning budget, a sampling strategy jointly balancing uncertainty and representativeness yielded the strongest effect. These findings provide compelling empirical evidence of the knowledge augmentation effect arising from human learning within a closed-loop active learning framework.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100775"},"PeriodicalIF":4.9,"publicationDate":"2025-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145417479","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Matthews correlation coefficient-based feature ranking in recursive ensemble feature selection for high-dimensional and low-sample size data 基于Matthews相关系数的高维低样本量数据递归集成特征选择中的特征排序

IF 4.9

Machine learning with applications

Pub Date : 2025-10-24 DOI: 10.1016/j.mlwa.2025.100757

David Rojas-Velazquez , Aletta D. Kraneveld , Alberto Tonda , Alejandro Lopez-Rincon

Identifying reliable biomarkers in omics data is challenging due to the high number of features and limited sample sizes, which often lead to overfitting, biased results, and poor reproducibility. These issues are further complicated by class imbalance, common in medical datasets. To address these challenges, we present MCC-REFS, an improved version of the Recursive Ensemble Feature Selection method. MCC-REFS uses the Matthews Correlation Coefficient (MCC) as a selection criterion, offering a more balanced evaluation of classification performance, especially in imbalanced datasets. Unlike traditional methods that require manual tuning or predefined feature counts, MCC-REFS automatically selects the most informative and compact feature sets using an ensemble of eight machine learning classifiers. We evaluated MCC-REFS on synthetic datasets and several real-world omics datasets, including mRNA expression profiles and multi-label breast cancer data. Compared to existing methods such as REFS, GRACES, DNP, and GCNN, MCC-REFS consistently achieved higher or comparable performance while selecting fewer features. Validation using independent classifiers confirmed the robustness of the selected features. Overall, MCC-REFS provides a scalable, flexible, and reliable approach for feature selection in biomedical research, with strong potential for diagnostic and prognostic applications.

由于大量的特征和有限的样本量，在组学数据中识别可靠的生物标志物是具有挑战性的，这通常会导致过拟合、有偏差的结果和较差的可重复性。在医疗数据集中常见的类别不平衡使这些问题进一步复杂化。为了解决这些挑战，我们提出了MCC-REFS，这是递归集成特征选择方法的改进版本。MCC- refs使用马修斯相关系数（MCC）作为选择标准，提供了更平衡的分类性能评估，特别是在不平衡的数据集。与需要手动调优或预定义特征计数的传统方法不同，MCC-REFS使用八个机器学习分类器的集合自动选择信息最丰富、最紧凑的特征集。我们在合成数据集和几个真实世界的组学数据集上评估了MCC-REFS，包括mRNA表达谱和多标签乳腺癌数据。与REFS、GRACES、DNP和GCNN等现有方法相比，MCC-REFS在选择更少特征的情况下始终实现更高或相当的性能。使用独立分类器的验证确认了所选特征的鲁棒性。总体而言，MCC-REFS为生物医学研究中的特征选择提供了一种可扩展、灵活和可靠的方法，在诊断和预后应用方面具有强大的潜力。

{"title":"Matthews correlation coefficient-based feature ranking in recursive ensemble feature selection for high-dimensional and low-sample size data","authors":"David Rojas-Velazquez , Aletta D. Kraneveld , Alberto Tonda , Alejandro Lopez-Rincon","doi":"10.1016/j.mlwa.2025.100757","DOIUrl":"10.1016/j.mlwa.2025.100757","url":null,"abstract":"<div><div>Identifying reliable biomarkers in omics data is challenging due to the high number of features and limited sample sizes, which often lead to overfitting, biased results, and poor reproducibility. These issues are further complicated by class imbalance, common in medical datasets. To address these challenges, we present MCC-REFS, an improved version of the Recursive Ensemble Feature Selection method. MCC-REFS uses the Matthews Correlation Coefficient (MCC) as a selection criterion, offering a more balanced evaluation of classification performance, especially in imbalanced datasets. Unlike traditional methods that require manual tuning or predefined feature counts, MCC-REFS automatically selects the most informative and compact feature sets using an ensemble of eight machine learning classifiers. We evaluated MCC-REFS on synthetic datasets and several real-world omics datasets, including mRNA expression profiles and multi-label breast cancer data. Compared to existing methods such as REFS, GRACES, DNP, and GCNN, MCC-REFS consistently achieved higher or comparable performance while selecting fewer features. Validation using independent classifiers confirmed the robustness of the selected features. Overall, MCC-REFS provides a scalable, flexible, and reliable approach for feature selection in biomedical research, with strong potential for diagnostic and prognostic applications.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100757"},"PeriodicalIF":4.9,"publicationDate":"2025-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145417478","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Unsupervised deep learning for semantic segmentation using laparoscopic videos: A self-detection and self-learning approach 使用腹腔镜视频进行语义分割的无监督深度学习：一种自我检测和自我学习方法

IF 4.9

Machine learning with applications

Pub Date : 2025-10-24 DOI: 10.1016/j.mlwa.2025.100767

Sina Saadati , Maryam Hashemi , Camran Nezhat

Artificial intelligence (AI) and machine learning methods play a crucial role in image processing applications, particularly in semantic segmentation and localization tasks. These models rely on annotated datasets to train algorithms capable of detecting and localizing objects of interest in images. However, the manual annotation process demands substantial human effort and focus, posing significant challenges in terms of time, economic costs, and energy consumption. This paper introduces a novel unsupervised deep learning approach inspired by the psychology of human learning to address these limitations. First, the self-learning methodology is proposed to utilize only one or two annotated images to train neural networks, enabling automated segmentation and annotation of a large volume of unannotated images within the dataset. Then, to enhance the automation of this process, a complementary object detection algorithm, termed Self-Detection, is proposed. By simply clicking on an object within an image, this algorithm differentiates it from other objects in the scene, streamlining object identification and segmentation. Integrating the proposed Self-Learning and Self-Detection methods results in a fully unsupervised framework for training semantic segmentation neural networks. The key outcomes of this methodology include (1) trained neural network models capable of precise segmentation and localization of objects of interest, and (2) a fully-automatically well-annotated image dataset suitable for training other types of AI models with diverse architectures. The proposed methodology can be used for developing accurate, reliable, and interpretable deep learning models for various tasks and applications, both medical and non-medical, as well as for segmentation or localization tasks.

人工智能（AI）和机器学习方法在图像处理应用中起着至关重要的作用，特别是在语义分割和定位任务中。这些模型依赖于带注释的数据集来训练能够检测和定位图像中感兴趣对象的算法。然而，手动注释过程需要大量的人力和精力，在时间、经济成本和能源消耗方面提出了重大挑战。本文介绍了一种受人类学习心理学启发的新型无监督深度学习方法来解决这些限制。首先，提出了利用一到两张标注图像训练神经网络的自学习方法，实现了对数据集中大量未标注图像的自动分割和标注。然后，为了提高这一过程的自动化程度，提出了一种互补的目标检测算法，称为自检测。通过简单地点击图像中的对象，该算法将其与场景中的其他对象区分开来，简化对象识别和分割。将所提出的自学习和自检测方法相结合，形成了一个完全无监督的语义分割神经网络训练框架。该方法的主要成果包括：(1)训练有素的神经网络模型能够精确分割和定位感兴趣的对象，以及(2)适合训练具有不同架构的其他类型人工智能模型的全自动良好注释图像数据集。所提出的方法可用于为各种任务和应用（包括医疗和非医疗）以及分割或定位任务开发准确、可靠和可解释的深度学习模型。

{"title":"Unsupervised deep learning for semantic segmentation using laparoscopic videos: A self-detection and self-learning approach","authors":"Sina Saadati , Maryam Hashemi , Camran Nezhat","doi":"10.1016/j.mlwa.2025.100767","DOIUrl":"10.1016/j.mlwa.2025.100767","url":null,"abstract":"<div><div>Artificial intelligence (AI) and machine learning methods play a crucial role in image processing applications, particularly in semantic segmentation and localization tasks. These models rely on annotated datasets to train algorithms capable of detecting and localizing objects of interest in images. However, the manual annotation process demands substantial human effort and focus, posing significant challenges in terms of time, economic costs, and energy consumption. This paper introduces a novel unsupervised deep learning approach inspired by the psychology of human learning to address these limitations. First, the self-learning methodology is proposed to utilize only one or two annotated images to train neural networks, enabling automated segmentation and annotation of a large volume of unannotated images within the dataset. Then, to enhance the automation of this process, a complementary object detection algorithm, termed Self-Detection, is proposed. By simply clicking on an object within an image, this algorithm differentiates it from other objects in the scene, streamlining object identification and segmentation. Integrating the proposed Self-Learning and Self-Detection methods results in a fully unsupervised framework for training semantic segmentation neural networks. The key outcomes of this methodology include (1) trained neural network models capable of precise segmentation and localization of objects of interest, and (2) a fully-automatically well-annotated image dataset suitable for training other types of AI models with diverse architectures. The proposed methodology can be used for developing accurate, reliable, and interpretable deep learning models for various tasks and applications, both medical and non-medical, as well as for segmentation or localization tasks.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100767"},"PeriodicalIF":4.9,"publicationDate":"2025-10-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145417481","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

On the retraining frequency of global models in retail demand forecasting 零售需求预测中全局模型的再训练频率研究

IF 4.9

Machine learning with applications

Pub Date : 2025-10-23 DOI: 10.1016/j.mlwa.2025.100769

Marco Zanotti

In an era of increasing computational capabilities and growing environmental consciousness, organizations face a critical challenge in balancing the accuracy of forecasting models with computational efficiency and sustainability. Global forecasting models, lowering the computational time, have gained significant attention over the years. However, the common practice of retraining these models with new observations raises important questions about the costs of forecasting. Using ten different machine learning and deep learning models, we analyzed various retraining scenarios, ranging from continuous updates to no retraining at all, across two large retail demand datasets. We showed that less frequent retraining strategies maintain the forecast accuracy while reducing the computational costs, providing a more sustainable approach to large-scale forecasting. We also found that machine learning models are a marginally better choice to reduce the costs of forecasting when coupled with less frequent model retraining strategies as the frequency of the data increases. Our findings challenge the conventional belief that frequent retraining is essential for maintaining forecasting accuracy. Instead, periodic retraining offers a good balance between predictive performance and efficiency, both in the case of point and probabilistic forecasting. These insights provide actionable guidelines for organizations seeking to optimize forecasting pipelines while reducing costs and energy consumption.

在计算能力不断增强和环境意识不断增强的时代，组织面临着平衡预测模型准确性与计算效率和可持续性的关键挑战。减少计算时间的全球预测模型多年来获得了极大的关注。然而，用新的观测结果对这些模型进行再训练的常见做法引发了有关预测成本的重要问题。使用十种不同的机器学习和深度学习模型，我们分析了两个大型零售需求数据集的各种再培训场景，从持续更新到根本不进行再培训。我们发现，较少频率的再训练策略在保持预测准确性的同时降低了计算成本，为大规模预测提供了一种更可持续的方法。我们还发现，随着数据频率的增加，机器学习模型是降低预测成本的一个稍微更好的选择，当加上不太频繁的模型再训练策略时。我们的发现挑战了传统观念，即频繁的再培训对于保持预测的准确性至关重要。相反，在点预测和概率预测的情况下，定期再训练提供了预测性能和效率之间的良好平衡。这些见解为寻求优化预测管道，同时降低成本和能源消耗的组织提供了可操作的指导方针。

{"title":"On the retraining frequency of global models in retail demand forecasting","authors":"Marco Zanotti","doi":"10.1016/j.mlwa.2025.100769","DOIUrl":"10.1016/j.mlwa.2025.100769","url":null,"abstract":"<div><div>In an era of increasing computational capabilities and growing environmental consciousness, organizations face a critical challenge in balancing the accuracy of forecasting models with computational efficiency and sustainability. Global forecasting models, lowering the computational time, have gained significant attention over the years. However, the common practice of retraining these models with new observations raises important questions about the costs of forecasting. Using ten different machine learning and deep learning models, we analyzed various retraining scenarios, ranging from continuous updates to no retraining at all, across two large retail demand datasets. We showed that less frequent retraining strategies maintain the forecast accuracy while reducing the computational costs, providing a more sustainable approach to large-scale forecasting. We also found that machine learning models are a marginally better choice to reduce the costs of forecasting when coupled with less frequent model retraining strategies as the frequency of the data increases. Our findings challenge the conventional belief that frequent retraining is essential for maintaining forecasting accuracy. Instead, periodic retraining offers a good balance between predictive performance and efficiency, both in the case of point and probabilistic forecasting. These insights provide actionable guidelines for organizations seeking to optimize forecasting pipelines while reducing costs and energy consumption.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100769"},"PeriodicalIF":4.9,"publicationDate":"2025-10-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363478","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

AI to protect AI: A modular pipeline for detecting label-flipping poisoning attacks AI保护AI：用于检测标签翻转中毒攻击的模块化管道

IF 4.9

Machine learning with applications

Pub Date : 2025-10-23 DOI: 10.1016/j.mlwa.2025.100768

Hossein Abroshan

Modern machine learning models are vulnerable to data poisoning attacks that compromise the integrity of their training data, with label flipping being a particularly insidious variant. In a label flipping attack, an adversary maliciously alters a fraction of the training labels to mislead the model, which can significantly degrade performance or cause targeted misclassifications while often evading simple detection. In this work, we address this threat by introducing a modular, attack-agnostic detection framework (“AI to Protect AI”) that monitors model behaviour for poisoning indicators without requiring internal access or changes to the target model. A Behaviour Monitoring Module (BMM) continuously observes the model’s outputs, extracting telltale features such as prediction probabilities, entropy, and margins for each input. These features are analysed by an ensemble of detector models, including supervised classifiers and unsupervised anomaly detectors, that collaboratively flag suspicious training samples indicative of label tampering. The proposed framework is dataset-agnostic and model-agnostic, as demonstrated across diverse image classification tasks using the MNIST (handwritten digits), CIFAR-10 (natural images), and ChestXray14 (medical X-rays) datasets. Experimental results indicate that the system reliably detects poisoned data with high accuracy (e.g., an area under the ROC curve exceeding 0.95 on MNIST, above 0.90 on CIFAR-10, and up to 0.85 on ChestXray14), while maintaining low false alarm rates. This work highlights a novel “AI to protect AI” approach, leveraging multiple lightweight detectors in concert to safeguard learning processes across different domains and thereby enhance the security and trustworthiness of AI systems.

现代机器学习模型很容易受到数据中毒攻击，这会损害其训练数据的完整性，其中标签翻转是一种特别阴险的变体。在标签翻转攻击中，攻击者恶意地改变训练标签的一小部分来误导模型，这可能会显著降低性能或导致目标错误分类，而通常会逃避简单的检测。在这项工作中，我们通过引入一个模块化的、攻击不可知的检测框架（“AI保护AI”）来解决这一威胁，该框架可以监控模型行为以获取中毒指标，而无需内部访问或更改目标模型。行为监测模块（BMM）持续观察模型的输出，提取出预测概率、熵和每个输入的边际等特征。这些特征通过检测器模型的集合进行分析，包括监督分类器和无监督异常检测器，它们协同标记指示标签篡改的可疑训练样本。所提出的框架与数据集和模型无关，正如使用MNIST（手写数字）、CIFAR-10（自然图像）和ChestXray14（医学x射线）数据集的不同图像分类任务所证明的那样。实验结果表明，该系统能够可靠、准确地检测出中毒数据（例如，在MNIST上，ROC曲线下面积超过0.95，在cifar10上超过0.90，在ChestXray14上高达0.85），同时保持较低的误报率。这项工作强调了一种新颖的“人工智能保护人工智能”方法，利用多个轻量级检测器协同保护不同领域的学习过程，从而增强人工智能系统的安全性和可信度。

{"title":"AI to protect AI: A modular pipeline for detecting label-flipping poisoning attacks","authors":"Hossein Abroshan","doi":"10.1016/j.mlwa.2025.100768","DOIUrl":"10.1016/j.mlwa.2025.100768","url":null,"abstract":"<div><div>Modern machine learning models are vulnerable to data poisoning attacks that compromise the integrity of their training data, with label flipping being a particularly insidious variant. In a label flipping attack, an adversary maliciously alters a fraction of the training labels to mislead the model, which can significantly degrade performance or cause targeted misclassifications while often evading simple detection. In this work, we address this threat by introducing a modular, attack-agnostic detection framework (“AI to Protect AI”) that monitors model behaviour for poisoning indicators without requiring internal access or changes to the target model. A Behaviour Monitoring Module (BMM) continuously observes the model’s outputs, extracting telltale features such as prediction probabilities, entropy, and margins for each input. These features are analysed by an ensemble of detector models, including supervised classifiers and unsupervised anomaly detectors, that collaboratively flag suspicious training samples indicative of label tampering. The proposed framework is dataset-agnostic and model-agnostic, as demonstrated across diverse image classification tasks using the MNIST (handwritten digits), CIFAR-10 (natural images), and ChestXray14 (medical X-rays) datasets. Experimental results indicate that the system reliably detects poisoned data with high accuracy (e.g., an area under the ROC curve exceeding 0.95 on MNIST, above 0.90 on CIFAR-10, and up to 0.85 on ChestXray14), while maintaining low false alarm rates. This work highlights a novel “AI to protect AI” approach, leveraging multiple lightweight detectors in concert to safeguard learning processes across different domains and thereby enhance the security and trustworthiness of AI systems.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100768"},"PeriodicalIF":4.9,"publicationDate":"2025-10-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363477","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Survey of neural network optimization methods for sustainable AI: From data preprocessing to hardware acceleration 可持续人工智能的神经网络优化方法综述：从数据预处理到硬件加速

IF 4.9

Machine learning with applications

Pub Date : 2025-10-21 DOI: 10.1016/j.mlwa.2025.100762

Omar Ghoneim, Petr Dobias, Olivier Romain

Technological developments in artificial intelligence and machine learning have recently been integrated into a range of daily objects to improve our lives. However, this progress has increased memory and energy consumption, causing harm to the environment. Addressing this challenge is critical for the well-being of current and future generations and for ensuring sustainability. The reduction of carbon footprint for neural networks is an essential step toward sustainable AI development. We present a structured examination of neural network optimization which covers the entire pipeline from data preprocessing to model design, compression, and hardware efficiency. Unlike prior works that focus on isolated stages, this survey integrates diverse strategies such as data labeling, feature selection, quantization, pruning, knowledge distillation, and approximate adders into a unified framework. This integration provides a cross-stage perspective that reveals synergies and trade-offs often hidden in fragmented studies, while also offering a consolidated reference for researchers through analysis and benchmarking. The novelty lies in combining a literature-wide synthesis with an interactive benchmarking platform that enables side-by-side comparison of optimization methods across metrics and deployment scenarios. A key contribution is the development of a platform compiling results from 139 peer-reviewed studies (81% from 2020 onward), enabling interactive exploration of accuracy, latency, and energy trade-offs. Validation comes from aggregated cross-study analysis, grounding insights in a broad and current evidence base rather than single experiments. This perspective is particularly valuable for guiding sustainable AI development by identifying trade-offs and synergies across optimization stages.

最近，人工智能和机器学习的技术发展已经融入到一系列日常用品中，以改善我们的生活。然而，这种进步增加了内存和能量消耗，对环境造成了危害。应对这一挑战对当代人和子孙后代的福祉以及确保可持续性至关重要。减少神经网络的碳足迹是实现人工智能可持续发展的重要一步。我们提出了神经网络优化的结构化检查，涵盖了从数据预处理到模型设计，压缩和硬件效率的整个管道。与以往的研究不同，该研究将数据标记、特征选择、量化、修剪、知识蒸馏和近似加法器等多种策略整合到一个统一的框架中。这种整合提供了一个跨阶段的视角，揭示了往往隐藏在碎片化研究中的协同效应和权衡，同时也通过分析和基准测试为研究人员提供了综合参考。其新颖之处在于将文献综合与交互式基准测试平台相结合，该平台可以跨度量和部署场景并排比较优化方法。其中一个关键贡献是开发了一个平台，汇集了139项同行评议研究的结果（从2020年起占81%），从而实现了对准确性、延迟和能源权衡的交互式探索。验证来自综合的交叉研究分析，基于广泛和当前的证据基础的见解，而不是单一的实验。这种观点对于通过确定优化阶段之间的权衡和协同作用来指导可持续的人工智能发展尤其有价值。

{"title":"Survey of neural network optimization methods for sustainable AI: From data preprocessing to hardware acceleration","authors":"Omar Ghoneim, Petr Dobias, Olivier Romain","doi":"10.1016/j.mlwa.2025.100762","DOIUrl":"10.1016/j.mlwa.2025.100762","url":null,"abstract":"<div><div>Technological developments in artificial intelligence and machine learning have recently been integrated into a range of daily objects to improve our lives. However, this progress has increased memory and energy consumption, causing harm to the environment. Addressing this challenge is critical for the well-being of current and future generations and for ensuring sustainability. The reduction of carbon footprint for neural networks is an essential step toward sustainable AI development. We present a structured examination of neural network optimization which covers the entire pipeline from data preprocessing to model design, compression, and hardware efficiency. Unlike prior works that focus on isolated stages, this survey integrates diverse strategies such as data labeling, feature selection, quantization, pruning, knowledge distillation, and approximate adders into a unified framework. This integration provides a cross-stage perspective that reveals synergies and trade-offs often hidden in fragmented studies, while also offering a consolidated reference for researchers through analysis and benchmarking. The novelty lies in combining a literature-wide synthesis with an interactive benchmarking platform that enables side-by-side comparison of optimization methods across metrics and deployment scenarios. A key contribution is the development of a platform compiling results from 139 peer-reviewed studies (81% from 2020 onward), enabling interactive exploration of accuracy, latency, and energy trade-offs. Validation comes from aggregated cross-study analysis, grounding insights in a broad and current evidence base rather than single experiments. This perspective is particularly valuable for guiding sustainable AI development by identifying trade-offs and synergies across optimization stages.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100762"},"PeriodicalIF":4.9,"publicationDate":"2025-10-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363481","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Machine learning techniques for analysing cardiotocography signals for early detection of fetal anomalies based on feature engineering methods 基于特征工程方法的早期检测胎儿异常的心脏造影信号分析机器学习技术

IF 4.9

Machine learning with applications

Pub Date : 2025-10-19 DOI: 10.1016/j.mlwa.2025.100766

Ibrahim Abunadi

Complications during childbirth are among the leading causes of infant mortality in the first months of life. Cardiotocography (CTG) is a primary tool for diagnosing and monitoring the condition of the fetus and identifying high-risk women during labor. Knowing the dynamic patterns of CTG signals requires a real-time interpretation. This is a difficult task, with significant differences in opinion between doctors. This study presents a new integrated framework that combines feature selection and dimensionality reduction methods to enhance classification of fetal health status from CTG signals. The primary contribution of this study is to combine the Shapley Additive Explanations (SHAP) method with both dimensionality reduction methods, t-distributed Stochastic Neighbor Embedding (t-SNE), and Principal Component Analysis (PCA). By implementing the combined approach, the dimensionality of the CTG data was reduced, while the most feasible features remained foundational for fetal health diagnosis. The proposed method was validated against the well-known, publicly available, and widely used CTG data. The use of a strong statistical method using the Variance Inflation Factor (VIF) facilitated a rigorous approach to reducing multicollinearity. The VIF value suggests that the data have improved quality and, therefore, reliability for the predictive model development. Various machine learning classifiers, including XGBoost, K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM), were trained and evaluated using low-dimensional feature sets generated via PCA and t-SNE. Among the different classifiers, the RF classifier achieved the best results with PCA features, achieving average results with an accuracy of 98 %, precision of 96.13 %, recall of 96.4 %, F1 score of 96.3 %, and AUC of 99.6 %.

分娩并发症是婴儿出生后最初几个月死亡的主要原因之一。心脏造影（CTG）是诊断和监测胎儿状况以及识别分娩过程中高危妇女的主要工具。了解CTG信号的动态模式需要实时解释。这是一项艰巨的任务，医生之间的意见存在显著差异。本研究提出了一种结合特征选择和降维方法的新的集成框架，以增强CTG信号对胎儿健康状况的分类。本研究的主要贡献是将Shapley加性解释（SHAP）方法与降维方法、t分布随机邻居嵌入（t-SNE）和主成分分析（PCA）相结合。通过实施联合方法，CTG数据的维数被降低，而最可行的特征仍然是胎儿健康诊断的基础。针对众所周知的、公开可用的、广泛使用的CTG数据，对所提出的方法进行了验证。使用方差膨胀因子（VIF）的强统计方法有助于减少多重共线性的严格方法。VIF值表明数据质量有所提高，因此，预测模型开发的可靠性更高。各种机器学习分类器，包括XGBoost， k -近邻（KNN），决策树（DT），随机森林（RF）和支持向量机（SVM），使用通过PCA和t-SNE生成的低维特征集进行训练和评估。在不同的分类器中，射频分类器在PCA特征上取得了最好的效果，平均准确率为98%，精密度为96.13%，召回率为96.4%，F1得分为96.3%，AUC为99.6%。

{"title":"Machine learning techniques for analysing cardiotocography signals for early detection of fetal anomalies based on feature engineering methods","authors":"Ibrahim Abunadi","doi":"10.1016/j.mlwa.2025.100766","DOIUrl":"10.1016/j.mlwa.2025.100766","url":null,"abstract":"<div><div>Complications during childbirth are among the leading causes of infant mortality in the first months of life. Cardiotocography (CTG) is a primary tool for diagnosing and monitoring the condition of the fetus and identifying high-risk women during labor. Knowing the dynamic patterns of CTG signals requires a real-time interpretation. This is a difficult task, with significant differences in opinion between doctors. This study presents a new integrated framework that combines feature selection and dimensionality reduction methods to enhance classification of fetal health status from CTG signals. The primary contribution of this study is to combine the Shapley Additive Explanations (SHAP) method with both dimensionality reduction methods, t-distributed Stochastic Neighbor Embedding (t-SNE), and Principal Component Analysis (PCA). By implementing the combined approach, the dimensionality of the CTG data was reduced, while the most feasible features remained foundational for fetal health diagnosis. The proposed method was validated against the well-known, publicly available, and widely used CTG data. The use of a strong statistical method using the Variance Inflation Factor (VIF) facilitated a rigorous approach to reducing multicollinearity. The VIF value suggests that the data have improved quality and, therefore, reliability for the predictive model development. Various machine learning classifiers, including XGBoost, K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM), were trained and evaluated using low-dimensional feature sets generated via PCA and t-SNE. Among the different classifiers, the RF classifier achieved the best results with PCA features, achieving average results with an accuracy of 98 %, precision of 96.13 %, recall of 96.4 %, F1 score of 96.3 %, and AUC of 99.6 %.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100766"},"PeriodicalIF":4.9,"publicationDate":"2025-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363476","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Toward AI-driven fire imagery: Attributes, challenges, comparisons, and the promise of VLMs and LLMs 走向人工智能驱动的火灾图像：vlm和llm的属性、挑战、比较和前景

IF 4.9

Machine learning with applications

Pub Date : 2025-10-18 DOI: 10.1016/j.mlwa.2025.100763

Sayed Pedram Haeri Boroujeni , Niloufar Mehrabi , Fatemeh Afghah , Connor Peter McGrath , Danish Bhatkar , Mithilesh Anil Biradar , Abolfazl Razi

Despite recent advancements in technology-driven fire management systems, environment continues to grapple with the increasing frequency and severity of wildfires, an issue exacerbated by climate change. A recent example is the devastating California wildfire in January 2025, which burned approximately 182,197 acres, destroyed 16,306 structures, and resulted in a total economic loss of over 275 billion dollars. Such events underscore the urgent need for further investment in intelligent, data-driven fire management solutions. One transformative development in this field has been the deployment of Unmanned Aerial Systems (UAS) for wildfire monitoring and control. These systems capture multimodal imagery and sensor data, facilitating the development of advanced Artificial Intelligence (AI) models for fire detection, spread modeling and prediction, effective suppression, and post-incident damage assessment. Unfortunately, most existing wildfire datasets exhibit significant heterogeneity in terms of imaging modalities (e.g., RGB, thermal, IR), annotation quality, target applications, and geospatial attributes. This diversity often complicates the identification of appropriate datasets for new and emerging wildfire scenarios, which remains a core challenge that hampers progress in the field and limits generalizability and reusability. This paper presents a comprehensive review of prominent wildfire datasets, offering a systematic comparison across various dimensions to help researchers, especially newcomers, select the most suitable datasets for their needs. Additionally, it identifies key parameters to consider when designing and collecting new fire imagery datasets to enhance future usability. Another key contribution of this work is its exploration of how emerging Large Language Models/Vision Language Models (LLMs/VLMs) can catalyze the creation, augmentation, and application of wildfire datasets. We discuss the potential of these models to integrate global knowledge for more accurate fire detection, devise evacuation plans, and support data-driven fire control strategies.

尽管技术驱动的火灾管理系统最近取得了进步，但环境仍在努力应对日益频繁和严重的野火，这一问题因气候变化而加剧。最近的一个例子是2025年1月毁灭性的加州野火，烧毁了大约182197英亩土地，摧毁了16306座建筑，造成了超过2750亿美元的经济损失。这些事件强调了对智能、数据驱动的火灾管理解决方案进行进一步投资的迫切需要。该领域的一个变革性发展是无人机系统（UAS）用于野火监测和控制的部署。这些系统捕获多模态图像和传感器数据，促进先进的人工智能（AI）模型的发展，用于火灾探测、蔓延建模和预测、有效抑制和事故后损害评估。不幸的是，大多数现有野火数据集在成像方式（如RGB、热、红外）、注释质量、目标应用和地理空间属性方面表现出显著的异质性。这种多样性往往使识别新的和正在出现的野火情景的适当数据集变得复杂，这仍然是阻碍该领域进展并限制通用性和可重用性的核心挑战。本文对突出的野火数据集进行了全面的回顾，提供了不同维度的系统比较，以帮助研究人员，特别是新手，选择最适合他们需要的数据集。此外，它还确定了设计和收集新的火灾图像数据集时需要考虑的关键参数，以增强未来的可用性。这项工作的另一个关键贡献是它探索了新兴的大型语言模型/视觉语言模型（llm / vlm）如何促进野火数据集的创建、增强和应用。我们讨论了这些模型在整合全球知识以实现更准确的火灾探测、制定疏散计划和支持数据驱动的火灾控制策略方面的潜力。

{"title":"Toward AI-driven fire imagery: Attributes, challenges, comparisons, and the promise of VLMs and LLMs","authors":"Sayed Pedram Haeri Boroujeni , Niloufar Mehrabi , Fatemeh Afghah , Connor Peter McGrath , Danish Bhatkar , Mithilesh Anil Biradar , Abolfazl Razi","doi":"10.1016/j.mlwa.2025.100763","DOIUrl":"10.1016/j.mlwa.2025.100763","url":null,"abstract":"<div><div>Despite recent advancements in technology-driven fire management systems, environment continues to grapple with the increasing frequency and severity of wildfires, an issue exacerbated by climate change. A recent example is the devastating California wildfire in January 2025, which burned approximately 182,197 acres, destroyed 16,306 structures, and resulted in a total economic loss of over 275 billion dollars. Such events underscore the urgent need for further investment in intelligent, data-driven fire management solutions. One transformative development in this field has been the deployment of Unmanned Aerial Systems (UAS) for wildfire monitoring and control. These systems capture multimodal imagery and sensor data, facilitating the development of advanced Artificial Intelligence (AI) models for fire detection, spread modeling and prediction, effective suppression, and post-incident damage assessment. Unfortunately, most existing wildfire datasets exhibit significant heterogeneity in terms of imaging modalities (e.g., RGB, thermal, IR), annotation quality, target applications, and geospatial attributes. This diversity often complicates the identification of appropriate datasets for new and emerging wildfire scenarios, which remains a core challenge that hampers progress in the field and limits generalizability and reusability. This paper presents a comprehensive review of prominent wildfire datasets, offering a systematic comparison across various dimensions to help researchers, especially newcomers, select the most suitable datasets for their needs. Additionally, it identifies key parameters to consider when designing and collecting new fire imagery datasets to enhance future usability. Another key contribution of this work is its exploration of how emerging Large Language Models/Vision Language Models (LLMs/VLMs) can catalyze the creation, augmentation, and application of wildfire datasets. We discuss the potential of these models to integrate global knowledge for more accurate fire detection, devise evacuation plans, and support data-driven fire control strategies.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100763"},"PeriodicalIF":4.9,"publicationDate":"2025-10-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363480","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Estimating vessel arrival times in global supply chains 估计全球供应链中的船舶到达时间

IF 4.9

Machine learning with applications

Pub Date : 2025-10-18 DOI: 10.1016/j.mlwa.2025.100751

Mathijs Pellemans , Jesper Slik , Sandjai Bhulai

The global maritime industry, which facilitates around 90% of the world’s trade, faces operational inefficiencies due to inconsistent Automatic Identification System (AIS) data and scalability challenges, limiting the reliability of current tracking systems. This study aims to improve the reliability of both Estimated Time of Arrival (ETA) and vessel destination predictions. We apply Machine Learning (ML) techniques to predict ETAs through both regression and classification models, while leveraging high-order Markov chains for destination prediction based on sequential maritime route patterns. To ensure the practical applicability of these models, we developed an end-to-end pipeline that incorporates waypoint-based trajectory compression, optimizing the handling of satellite-based AIS data (s-AIS). Using a real-world dataset of global shipping routes, our ML models, particularly Support Vector Regression (SVR), achieved a lower mean absolute error than captain-provided ETAs (16.01 vs. 22.15 h). When the time to arrival was more than 75 h, SVR outperformed captain-provided ETAs, whereas captain-provided ETAs were more accurate in the final three days before arrival, due to frequent manual updates. High-order Markov chains achieved a near-perfect accuracy of 99.00% (std: 1.19%) in destination prediction, confirming the regularity of cargo ship routes. These findings demonstrate the potential of combining ML models with Markov chains to enhance the accuracy and reliability of long-term maritime logistics forecasting, transforming raw s-AIS data into actionable insights for improved operational decision-making.

由于自动识别系统（AIS）数据不一致和可扩展性挑战，限制了当前跟踪系统的可靠性，促进了全球约90%的贸易的全球海运业面临着运营效率低下的问题。本研究旨在提高预计到达时间（ETA）和船舶目的地预测的可靠性。我们应用机器学习（ML）技术通过回归和分类模型来预测eta，同时利用高阶马尔可夫链进行基于顺序海上航线模式的目的地预测。为了确保这些模型的实际适用性，我们开发了一个端到端管道，其中包含基于航路点的轨迹压缩，优化了基于卫星的AIS数据（s-AIS）的处理。使用全球航线的真实数据集，我们的ML模型，特别是支持向量回归（SVR），比船长提供的ETAs （16.01 vs 22.15 h）实现了更低的平均绝对误差。当到达时间超过75小时时，SVR优于机长提供的eta，而由于频繁的手动更新，机长提供的eta在到达前的最后三天更为准确。高阶马尔可夫链在目的地预测上达到了近乎完美的99.00%（标准差：1.19%），证实了货船航线的规律性。这些发现表明，将ML模型与马尔可夫链相结合，可以提高长期海上物流预测的准确性和可靠性，将原始的s-AIS数据转化为可操作的见解，从而改善运营决策。

{"title":"Estimating vessel arrival times in global supply chains","authors":"Mathijs Pellemans , Jesper Slik , Sandjai Bhulai","doi":"10.1016/j.mlwa.2025.100751","DOIUrl":"10.1016/j.mlwa.2025.100751","url":null,"abstract":"<div><div>The global maritime industry, which facilitates around 90% of the world’s trade, faces operational inefficiencies due to inconsistent Automatic Identification System (AIS) data and scalability challenges, limiting the reliability of current tracking systems. This study aims to improve the reliability of both Estimated Time of Arrival (ETA) and vessel destination predictions. We apply Machine Learning (ML) techniques to predict ETAs through both regression and classification models, while leveraging high-order Markov chains for destination prediction based on sequential maritime route patterns. To ensure the practical applicability of these models, we developed an end-to-end pipeline that incorporates waypoint-based trajectory compression, optimizing the handling of satellite-based AIS data (s-AIS). Using a real-world dataset of global shipping routes, our ML models, particularly Support Vector Regression (SVR), achieved a lower mean absolute error than captain-provided ETAs (16.01 vs. 22.15 h). When the time to arrival was more than 75 h, SVR outperformed captain-provided ETAs, whereas captain-provided ETAs were more accurate in the final three days before arrival, due to frequent manual updates. High-order Markov chains achieved a near-perfect accuracy of 99.00% (std: 1.19%) in destination prediction, confirming the regularity of cargo ship routes. These findings demonstrate the potential of combining ML models with Markov chains to enhance the accuracy and reliability of long-term maritime logistics forecasting, transforming raw s-AIS data into actionable insights for improved operational decision-making.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100751"},"PeriodicalIF":4.9,"publicationDate":"2025-10-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363592","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A machine learning approach to vulnerability detection combining software metrics and topic modelling: Evidence from smart contracts 结合软件度量和主题建模的漏洞检测机器学习方法：来自智能合约的证据

IF 4.9

Machine learning with applications

Pub Date : 2025-10-17 DOI: 10.1016/j.mlwa.2025.100759

Giacomo Ibba , Rumyana Neykova , Marco Ortu , Roberto Tonelli , Steve Counsell , Giuseppe Destefanis

This paper introduces a methodology for software vulnerability detection that combines structural and semantic analysis through software metrics and topic modelling. We evaluate the approach using smart contracts as a case study, focusing on their structural properties and the presence of known security vulnerabilities. We identify the most relevant metrics for vulnerability detection, evaluate multiple machine learning classifiers for both binary and multi-label classification, and improve classification performance by integrating topic modelling techniques.

Our analysis shows that metrics such as cyclomatic complexity, nesting depth, and function calls are strongly associated with vulnerability presence. Using these metrics, the Random Forest classifier achieved strong performance in binary classification (AUC: 0.982, accuracy: 0.977, F1-score: 0.808) and multi-label classification (AUC: 0.951, accuracy: 0.729, F1-score: 0.839). The addition of topic modelling using Non-Negative Matrix Factorisation further improved results, increasing the F1-score to 0.881. The evaluation is conducted on Ethereum smart contracts written in Solidity.

本文介绍了一种通过软件度量和主题建模，将结构分析和语义分析相结合的软件漏洞检测方法。我们使用智能合约作为案例研究来评估这种方法，重点关注它们的结构属性和已知安全漏洞的存在。我们确定了最相关的漏洞检测指标，评估了二元和多标签分类的多个机器学习分类器，并通过集成主题建模技术提高了分类性能。我们的分析表明，圈复杂度、嵌套深度和函数调用等指标与漏洞存在密切相关。使用这些指标，Random Forest分类器在二元分类（AUC: 0.982，准确率：0.977,F1-score: 0.808）和多标签分类（AUC: 0.951，准确率：0.729,F1-score: 0.839）方面取得了较好的表现。使用非负矩阵分解添加主题建模进一步改善了结果，将f1得分提高到0.881。评估是对用Solidity编写的以太坊智能合约进行的。

{"title":"A machine learning approach to vulnerability detection combining software metrics and topic modelling: Evidence from smart contracts","authors":"Giacomo Ibba , Rumyana Neykova , Marco Ortu , Roberto Tonelli , Steve Counsell , Giuseppe Destefanis","doi":"10.1016/j.mlwa.2025.100759","DOIUrl":"10.1016/j.mlwa.2025.100759","url":null,"abstract":"<div><div>This paper introduces a methodology for software vulnerability detection that combines structural and semantic analysis through software metrics and topic modelling. We evaluate the approach using smart contracts as a case study, focusing on their structural properties and the presence of known security vulnerabilities. We identify the most relevant metrics for vulnerability detection, evaluate multiple machine learning classifiers for both binary and multi-label classification, and improve classification performance by integrating topic modelling techniques.</div><div>Our analysis shows that metrics such as cyclomatic complexity, nesting depth, and function calls are strongly associated with vulnerability presence. Using these metrics, the Random Forest classifier achieved strong performance in binary classification (AUC: 0.982, accuracy: 0.977, F1-score: 0.808) and multi-label classification (AUC: 0.951, accuracy: 0.729, F1-score: 0.839). The addition of topic modelling using Non-Negative Matrix Factorisation further improved results, increasing the F1-score to 0.881. The evaluation is conducted on Ethereum smart contracts written in Solidity.</div></div>","PeriodicalId":74093,"journal":{"name":"Machine learning with applications","volume":"22 ","pages":"Article 100759"},"PeriodicalIF":4.9,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145363475","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0