Expert Systems with Applications最新文献

{"title":"Label-wise reliability-aware classifier for robust chest X-ray multi-label classification","authors":"Wenkai Ye,&nbsp;Xichen Ye,&nbsp;Hang Yao,&nbsp;Kequan Yang,&nbsp;Xiaoqiang Li","doi":"10.1016/j.eswa.2026.131438","DOIUrl":"10.1016/j.eswa.2026.131438","url":null,"abstract":"<div><div>Chest X-ray (CXR) multi-label classifiers are commonly trained with labels extracted from clinical reports, which are often incomplete and noisy. Under such label noise, we observe that performance degrades severely on tail classes (e.g., rare diseases), because these categories are under-represented and easily overwhelmed by corrupted annotations. As a result, existing methods can misidentify tail classes as noise and downweight their contribution to optimization during training. To address this issue, we propose LRC-CXR (<strong>L</strong>abel-wise <strong>R</strong>eliability-aware <strong>C</strong>lassifier for Chest X-ray), which calculates per-label reliability and selectively corrects noisy labels, preventing tail classes from being inadvertently under-trained. First, a Medical Description Bank provides lesion-aware textual prompts that guide the visual encoder toward diagnostically relevant patterns. Second, LRC-CXR models per-label reliability with a two-component Gaussian Mixture Model to distinguish clean, inseparable, and noisy labels. Third, only labels identified as noisy are refined via feature-space k-nearest-neighbor smoothing, while clean and inseparable labels are trained with stronger objectives through a hierarchical loss. Experiments on ChestX-ray14, CheXpert, and PadChest, including high-noise stress tests, show that LRC-CXR improves overall AUC/F1 and substantially boosts tail-class recall and robustness under label noise.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131438"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192694","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"An efficient dominance decomposition-based deep graph evolutionary algorithm for the expensive multi-objective optimization","authors":"Xing Cai ,&nbsp;Tong Zhang ,&nbsp;Zhen Cui","doi":"10.1016/j.eswa.2026.131379","DOIUrl":"10.1016/j.eswa.2026.131379","url":null,"abstract":"<div><div>To efficiently solve expensive multi-objective optimization problems (EMOPs), it is essential to identify valuable evaluation points that lead to optimal solutions with minimal computational cost. In this work, we propose a deep graph-based evolutionary algorithm, named the multi-objective evolutionary algorithm based on dominance decomposition and graph neural networks (MOEA-DDG). To model the complex dominance relationships among candidate solutions, an adjacency graph is constructed that integrates both evaluated and unevaluated individuals. A collaborative surrogate framework based on graph neural networks is proposed to guide the selection of promising candidates. This framework comprises two specialized models: the relational model (R-model), which decomposes dominance prediction into simpler sub-tasks by comparing solution quality across individual objectives-thus improving robustness and accuracy; and the metric model (M-model), which estimates solution quality by predicting Hypervolume (HV) improvement, enabling effective ranking when objective values are unavailable. To ensure thorough exploration of the solution space, a cluster-based selection strategy is designed, which partitions the objective domain and selects representative candidates from each cluster during each iteration. Extensive experiments on two benchmark test suites and a real-world molecular design task demonstrate that MOEA-DDG achieves a strong balance between exploration and exploitation, and significantly outperforms state-of-the-art algorithms under limited evaluation budgets.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131379"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192695","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Select prompting with chain-of-thought paired with large language models","authors":"Xun Che ,&nbsp;Wenjia Wu ,&nbsp;Yadang Chen ,&nbsp;Luanjuan Jiang ,&nbsp;Qianmu Li","doi":"10.1016/j.eswa.2026.131511","DOIUrl":"10.1016/j.eswa.2026.131511","url":null,"abstract":"<div><div>Chain-of-Thought (CoT) prompting has been demonstrated as a powerful tool for enhancing logical reasoning abilities. Among various methods, automatic prompting approaches like Auto-CoT have gained significant attention from researchers. However, existing methods often generate prompts with weak relevance to downstream tasks and overly simplistic content, which hinders the effectiveness of Large Language Models (LLMs) in addressing complex reasoning tasks. To address these limitations, we propose Select-Prompt, a novel automated prompting approach. It includes two key components: (1) an adaptive method for mining difficult samples, which improves task relevance, and (2) a reasoning chain selection strategy, which enhances prompt diversity through answer validation and gradient optimization. The proposed method has been thoroughly tested and validated on six reasoning datasets, encompassing arithmetic, commonsense, and symbolic reasoning tasks. Experimental results demonstrate that Select-Prompt outperforms state-of-the-art methods such as Auto-CoT and Self-Refine, significantly enhancing the accuracy and robustness of LLMs when reasoning through complex tasks.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131511"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192697","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Text-prompted generative data augmentation and semi-supervised learning for indoor defect detection using quadruped robots","authors":"Xiang Ji ,&nbsp;Junjie Chen ,&nbsp;Yonglin Fu ,&nbsp;Isabelle Chan ,&nbsp;Zhen Dong","doi":"10.1016/j.eswa.2026.131545","DOIUrl":"10.1016/j.eswa.2026.131545","url":null,"abstract":"<div><div>Timely defect detection is the key to formulating targeted maintenance plans to extend a facility’s lifespan. While tremendous efforts have been made in deploying robots (e.g., drones) for outdoor defect detection, little attention has been paid to defects taking place indoors. Indoor defect detection (IDD) has distinctive characteristics concerning (a) the complex environment (narrow passages, staircases, etc.) that challenges inspection data collection, and (b) the drastic image feature variation caused by uneven illumination and view point changes, which renders methods viable for outdoor detection less useful. This research takes on the challenges and proposes an automated IDD approach. To navigate challenging indoor environments (e.g., staircases), a quadruped robot platform is proposed for inspection image collection. To address the scarcity of indoor data, a novel algorithmic framework for IDD is formulated that integrates large generative models for data augmentation and semi-supervised learning to train on the generated unlabeled data. It is found that the proposed approach can effectively inspect challenging indoor space for defect detection by leveraging the unique locomotion capability of legged robots. Despite the lack of training data, the framework resulted in a performance gain of 5.03% for the model. Future research is suggested to explore autonomous navigation of the robots and three dimensional modeling of the detected defects.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131545"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192849","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Shipper-Centric Short-Term Vehicle Capacity Requirement Planning for First-Mile Bulk-Commodity Delivery","authors":"Jin-Myeong Jang,&nbsp;Hwa-Joong Kim","doi":"10.1016/j.eswa.2026.131531","DOIUrl":"10.1016/j.eswa.2026.131531","url":null,"abstract":"<div><div>This study considers the short-term vehicle capacity requirement planning problem from the shipper’s perspective, driven by the rapidly increasing demand for transporting bulk commodities in first mile logistics. The problem is to determine the types and quantities of vehicles and assign them to regions. The objective is to minimize the total costs associated with transport and additional stop costs. This problem is formulated as a mixed-integer programming model to reflect practical logistics constraints. This proposes two matheuristic algorithms to solve this problem. The relax-and-fix heuristic constructs an initial solution by progressively relaxing and fixing subsets of decision variables, while the fix-and-optimize heuristic improves this solution through iterative local optimization, refixing certain variables. Based on synthetic and real data from a South Korean context, computational experiments show that these proposed algorithms consistently produce acceptable quality solutions within a reasonable computation time. Scenario analyses provide further practical insights for shippers by incorporating real-world operational constraints.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131531"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192785","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Frequency-domain modulated spatio-temporal graph convolutional network for traffic flow prediction","authors":"Mengchao Liu,&nbsp;Bingchuan Jiang,&nbsp;Jingxu Liu,&nbsp;Ke Li","doi":"10.1016/j.eswa.2026.131474","DOIUrl":"10.1016/j.eswa.2026.131474","url":null,"abstract":"<div><div>Accurate traffic flow forecasting plays a pivotal role in relieving congestion and enhancing the efficiency of intelligent transportation systems. However, this task remains challenging given the coexistence of long-term periodic trends, short-term fluctuations, and highly dynamic spatial correlations in road networks. Most existing studies rely heavily on time-domain modeling, which limits their ability to capture multi-scale temporal dynamics and adaptive spatial dependencies. To address these issues, we propose a frequency-domain modulated spatio-temporal-graph convolutional network (FMTGCN), a collaborative framework that integrates frequency-domain modulation with adaptive spatio-temporal learning. More precisely, FMTGCN incorporates a joint temporal-positional encoding module, aiming to alleviate spatio-temporal confusion. It also features a frequency-domain modeling unit equipped with learnable amplitude and phase modulation for the purpose of capturing long-term dependencies. Additionally, a spectrum-driven dynamic graph convolution mechanism is introduced, which adaptively constructs spatial topology based on spectral features. To complement global modeling, temporal convolution layers are employed to capture local variations. Extensive experiments on four real-world traffic datasets demonstrate that FMTGCN consistently outperforms state-of-the-art baselines in both prediction accuracy and computational efficiency.Notably, on the PEMS08 dataset, FMTGCN improved training speed by 57.17% compared to the second-best method; in addition, it reduced the Mean Absolute Error (MAE) by a further 2.8%, validating its effectiveness and scalability for large-scale traffic forecasting.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131474"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192589","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 two-stage multimodal learning framework based on text-driven vision pretraining and cross-modal feature fusion for thyroid ultrasound diagnosis","authors":"Mengzhu Yu ,&nbsp;Yu Yan ,&nbsp;Tianwei Yan ,&nbsp;Zihan Xi ,&nbsp;Junchao Zeng ,&nbsp;Tao Huang ,&nbsp;Ming Xu ,&nbsp;Mingyue Ding","doi":"10.1016/j.eswa.2026.131440","DOIUrl":"10.1016/j.eswa.2026.131440","url":null,"abstract":"<div><div>Thyroid cancer is the most common endocrine cancer with a steadily rising incidence. In recent years, artificial intelligence, particularly deep learning, has shown great potential in assisting with the diagnosis of thyroid nodules. Despite these advances, existing approaches still face several challenges, including limited single-modal information, inadequate multimodal feature fusion, and a scarcity of multimodal datasets. Hence, this study proposes a Two-stage Text-driven Multimodal Network (TTM-Net), a novel network designed for thyroid nodule ultrasound diagnosis. In the first stage, a Text-Driven Vision Pretraining (TDVP) strategy is introduced to align visual and textual features using large-scale unlabeled text-image pairs, thereby enhancing the visual encoder’s ability to capture clinically relevant patterns from the thyroid ultrasound images. In the second stage, supervised classification training is conducted using labeled data. Visual and textual features are integrated via the Cross-Modal Concat-Attention Fusion (CMCAF) module, and a joint optimization strategy is employed to improve diagnostic performance. To overcome the limitations of TI RADS, we propose a Thyroid Nodule Ultrasound Feature Classification System for Artificial Intelligence (TNUFCS-AI), which integrates clinical domain knowledge. Based on this system, we constructed a structured multimodal thyroid dataset comprising 22,544 ultrasound images and corresponding clinical reports from 9466 patients. Experiments show that TTM-Net outperforms SOTA methods in classification accuracy, stability, and interpretability, achieving (94.98 ± 0.15%) accuracy. In summary, TTM-Net significantly improves the diagnostic performance and increases the model robustness while alleviating the dependence on labeled data, providing a viable solution for developing high-performance, interpretable, and deployable AI-assisted diagnostic systems for thyroid nodules.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131440"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192584","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Explicitly learning semantic relevance for salient object detection in remote sensing images","authors":"Tao Gao ,&nbsp;Weiguang Zhao ,&nbsp;Mengkun Liu ,&nbsp;Ting Chen ,&nbsp;Ziqi Li","doi":"10.1016/j.eswa.2026.131387","DOIUrl":"10.1016/j.eswa.2026.131387","url":null,"abstract":"<div><div>Salient object detection in remote sensing images (RSI-SOD) is crucial for computer vision in both high-altitude and low-altitude scenarios. Most existing methods primarily focus on multiscale feature integration, yet they encounter difficulties in achieving precise segmentation, particularly when confronted with complex object topologies and cluttered backgrounds. To address this, we propose a novel framework, ELSRNet, tailored to capturing the intrinsic semantic differences among features with diverse attributes, thereby facilitating pixel-wise separation of salient regions. This approach incorporates the deployment of a Foreground-Background Semantic Perception module (FBSP), which explicitly scrutinizes the semantic interactions through a more comprehensive Attention Guided Loss, ultimately strengthening the capacity to learn objects with complex structural characteristics. Going further, considering that the coupling between noise norms and convolutional kernels in cluttered backgrounds may amplify irrelevant responses and lead to false saliency predictions, the Non-Matching Feature Enhancement block (NMFE) is introduced to suppress such interference based on matching scores, and further refine the features through a gating mechanism. Concluding the process, the Global Perceptual Feature Aggregation module (GPFA) is designed to decouple features into semantic and structural information. It achieves saliency region localization while preserving fine-grained boundaries, producing high-quality saliency detection results. Experimental results and theoretical analysis reveal that the proposed network outperforms existing methods in enhancing detection capabilities across three benchmark datasets.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131387"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192710","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 deep learning-based multi-modal approach to robust place recognition in challenging orchard environments","authors":"Dilshan Ranasinghe,&nbsp;Chao Chen","doi":"10.1016/j.eswa.2026.131543","DOIUrl":"10.1016/j.eswa.2026.131543","url":null,"abstract":"<div><div>Place recognition addresses the problem of identifying the current location on revisit using the acquired knowledge of all the previously explored locations. The abstract representations of the locations can be generated as they are being explored using processed single or multi-sensory inputs. Later, the current abstraction of the environment can be compared with previously collected information to detect revisits. The loop closure component of Simultaneous Localization And Mapping (SLAM) uses these techniques to identify revisits, which allow the SLAM system to mitigate the associated long-term drift. RGB images, LiDAR point clouds, and LiDAR intensities are the three most popular sensory inputs used in place recognition literature. However, they each have certain disadvantages when used as a single-modal input. Additionally, in recent years, Deep Neural Network (DNN) based methods have emerged increasingly in the literature that addresses the place recognition problem. Therefore, In this work, we introduce a novel multi-modal method that takes advantage of the rich complementary information provided by the above three modalities along with a DNN for place recognition. This work was evaluated on multiple publicly available datasets as well as on a highly repetitive orchard dataset collected by our team. The results demonstrate the ability of this method to be used even in highly challenging environments such as orchards.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131543"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192700","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Class-balanced OpenMax for open-set recognition with long-tail sonar images","authors":"Jie Li ,&nbsp;Wenpei Jiao ,&nbsp;Jianlei Zhang ,&nbsp;Chunyan Zhang ,&nbsp;Guangming Xie","doi":"10.1016/j.eswa.2026.131463","DOIUrl":"10.1016/j.eswa.2026.131463","url":null,"abstract":"<div><div>Open-set recognition (OSR) of sonar images in real underwater scenarios faces compounded challenges arising from the interplay of long-tailed distributions and limited sample availability. Although open-set recognition theory has been widely integrated into advanced recognition frameworks, its fundamental module, OpenMax, still exhibits critical limitations under long-tailed scenarios: the use of a fixed tail size for extreme value theory (EVT) fitting often leads to unstable Weibull modeling for tail classes with scarce samples, thereby inducing systematic biases in the decision boundaries for unknown classes. To address this limitation, we propose a class-balanced dynamic tail size strategy that adaptively adjusts the number of tail samples for different categories, enabling extreme value modeling to account for both head and tail class statistics and thereby substantially improving its robustness and decision reliability. Extensive experiments conducted on four sonar image datasets from diverse sensors and scenarios demonstrate that the proposed strategy brings effective performance gains to OpenMax within the scope of our evaluation settings. The resulting Class-Balanced OpenMax can be integrated into advanced open-set recognition frameworks, achieving consistent improvements in both open-set and closed-set recognition of sonar images. Furthermore, this work provides empirical evidence and methodological insights for sonar open long-tailed recognition, contributing to the development of robust perception capabilities for underwater unmanned systems in complex environments.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"312 ","pages":"Article 131463"},"PeriodicalIF":7.5,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192842","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}