Applied Intelligence最新文献

{"title":"Efficient few-shot visual anomaly detection and localization via prompt tuning","authors":"Jiwoo Choi,&nbsp;Chang Ouk Kim","doi":"10.1007/s10489-026-07497-3","DOIUrl":"10.1007/s10489-026-07497-3","url":null,"abstract":"<div><p>Visual anomaly detection is crucial for detecting defective products in the manufacturing industry. The traditional anomaly detection models rely on convolutional neural networks that are pre-trained on datasets such as ImageNet. Recently, the vision-language model (VLM) CLIP has gained attention due to its ability to learn diverse visual concepts and its remarkable zero-/few-shot performance. In this paper, we aim to leverage the contextual understanding capabilities of CLIP in anomaly detection tasks. Adapting CLIP to various downstream tasks requires prompt engineering, where labels are transformed into a natural language form called “prompts” for vision-language alignment. However, manual prompt engineering requires domain-specific knowledge and encounters challenges when adapting to dynamic environments. To address these issues, we propose a few-shot anomaly prompt tuning method, which enables prompts to be optimized through training. Furthermore, to reduce false positives in anomaly maps under the few-shot regime, we introduce a top-<i>k</i> anomaly score ensemble (TASE) to mitigate model overconfidence. The proposed model achieves AUROCs of 95.6% and 88.5% on the MVTec-AD and VisA datasets, respectively, under a 4-shot setting, with a 92% inference speed improvement over a prior method, highlighting its effectiveness and efficiency.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035256","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Optimization of technical analysis-driven algorithmic trading using deep reinforcement learning","authors":"Yiqing Wang,&nbsp;Xianchang Wang,&nbsp;Xiaodong Liu","doi":"10.1007/s10489-026-07478-6","DOIUrl":"10.1007/s10489-026-07478-6","url":null,"abstract":"<div><p>Technical analysis indicators are critical for trend prediction in algorithmic trading. However, relying on a single technical indicator strategy cannot obtain consistent profits in dynamic markets. To address this limitation, we develop a dual-agent adaptive trading framework, and combine the double deep Q-network (DDQN) with three classic technical indicators (relative strength index (RSI), Williams percent range (WR), and commodity channel index (CCI)) respectively to construct three distinct base models (RSI-DDQN, WR-DDQN, CCI-DDQN). The framework classifies market data into different market strengths by comparing indicator values with a preset neutral threshold. For each base model, two specialized agents are developed for strong and weak market data. These agents optimize their trading strategies by interacting with environments constructed from the corresponding classified data. The framework dynamically selects the most suitable agent’s decisions as trading signals, based on market strength assessments. To further improve decision reliability, an integrated model is introduced by combining signals from the three base models through a hard voting mechanism. Empirical results from historical data of Devon Energy (DVN), Tesla (TSLA), NVIDIA (NVDA), and Apple (AAPL) show the framework’s effectiveness, yielding average Sharpe ratios of 1.06, 0.46, 0.48, and 0.81 for the RSI-DDQN, WR-DDQN, CCI-DDQN, and integrated models, respectively.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035451","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MvCConvFormer: The graph convolution transformer network with self-distillation for multi-view graph clustering","authors":"Jinke Wang,&nbsp;Xuefang Zhang,&nbsp;Chunchun Zhang,&nbsp;Xiangyang Yang,&nbsp;Yaliang Zhao","doi":"10.1007/s10489-026-07431-7","DOIUrl":"10.1007/s10489-026-07431-7","url":null,"abstract":"<div><p>Multi-view graph clustering has attracted increasing attention in recent research, which is widely applied in fields such as medicine, computer vision, and environmental monitoring. Although a rich line of works of multi-view graph clustering has achieved remarkable performance, some challenges still exist:1) Most of the studies either adopt shallow models or simple models with deep learning such as graph convolution networks (GCN). The core neighborhood aggregation operation of GCNs essentially performs weighted averaging over local features, which indiscriminately mixes valid information and noise during the aggregation process. Consequently, the representation capacity of such methods is severely limited when processing large-scale data with complex characteristics, including node-level topological heterogeneity and high inter-view noise; 2) Existing multi-view fusion approaches fail to fully exploit the complementary information across different views. With these issues in mind, we propose an innovative method for multi-view graph clustering named the graph convolution Transformer network with self-distillation (MvCConvFormer). Our novelty lies in three aspects. First, we combine the pre-training model and training model based on Transformer architecture to jointly learn data representation. Among the pre-training model, we adopt a new model with message passing operation like GCN with shallow layers to capture node-level feature and structural information from each view. Second, we apply the Transformer architecture to effectively fuse the structural information of the optimal adjacency matrix and the node representations learned by the pre-training model, thereby mitigating noise interference across multiple views. Third, self-distillation is utilized to improve the model generalization ability. Extensive experiments on various benchmark datasets verify the superiority of our method compared to the state-of-the-art multi-view graph clustering methods. Thus, we believe that good results can be achieved in practical scenarios, such as Environmental Monitoring.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035119","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FS-GSR: Graph-based scoring and sparse reconstruction for group streaming feature selection","authors":"Tianyuan Jia","doi":"10.1007/s10489-026-07498-2","DOIUrl":"10.1007/s10489-026-07498-2","url":null,"abstract":"<div><p>High-dimensional data often presents the core challenge of features arriving dynamically in grouped streams. This paper proposes a novel feature selection algorithm named Feature Selection with Group Streaming via Graph-based Scoring and Sparse Reconstruction (FS-GSR). The method establishes a three-stage framework comprising graph-guided hybrid filtering, intra-group structure preservation, and inter-group sparse reconstruction. This framework effectively leverages both the geometric structure and discriminative information of the data, ensuring high classification performance while significantly eliminating feature redundancy. Numerical experiments on multiple real-world datasets demonstrate that FS-GSR outperforms existing state-of-the-art methods in terms of both classification accuracy and feature compression capability, exhibiting superior overall performance and robustness.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://link.springer.com/content/pdf/10.1007/s10489-026-07498-2.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035288","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"POCS: phased optimal client selection for federated learning","authors":"Xuerui Li,&nbsp;Yangming Zhao,&nbsp;Chunming Qiao","doi":"10.1007/s10489-026-07462-0","DOIUrl":"10.1007/s10489-026-07462-0","url":null,"abstract":"<div><p>Traditional federated learning suffers from excessive communication overhead due to iterative model exchanges, long training time caused by data heterogeneity, and vulnerability to adversarial attacks due to gradient leakage risks. But if one strategically selects some specific clients for training during different training rounds, one can reduce the costs and the training time, while defending against the adversarial attacks. In this work, we propose the Phased Optimal Client Selection for Federated Learning (POCS) method, which includes customized optimal Client Selection method for each specific training phase in federated learning. We evaluate POCS’s performance with that of five state-of-the-art baselines, while the results indicate that POCS can improve the accuracy, reduce the training time and the communication costs, and show better defense against the membership inference attacks.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://link.springer.com/content/pdf/10.1007/s10489-026-07462-0.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035438","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"HALSI: Hierarchical adaptive latent skill inference for temporal abstraction in reinforcement learning","authors":"Xiaxi Si,&nbsp;Yang Ping,&nbsp;Tao Wang,&nbsp;Kun Chen,&nbsp;Yue Gao","doi":"10.1007/s10489-026-07485-7","DOIUrl":"10.1007/s10489-026-07485-7","url":null,"abstract":"<div><p>Learning temporally abstracted policies is challenging in long-horizon, sparse-reward reinforcement learning. Skill-based hierarchical frameworks alleviate this by decomposing tasks into reusable behavior primitives, but existing methods often fix skill lengths and use oversimplified latent representations, limiting temporal flexibility and behavioral diversity. We propose <b>HALSI</b>—Hierarchical Adaptive Latent Skill Inference, an end-to-end framework that jointly learns which skill to execute, for how long, and how to adjust them online. HALSI infers global latent skills from variable-length trajectories via a CLS-token Transformer encoder and generates coherent action sequences with a conditional diffusion policy. To overcome fixed horizons, we introduce an adaptive duration policy that predicts discrete execution lengths per skill. A multi-scale PPO scheme integrates a high-level modulation policy, refining latent skills with task-aware residuals, and a low-level soft-blending controller for step-wise corrections. Experiments on Fetch and D4RL Kitchen benchmarks show HALSI surpasses state-of-the-art hierarchical RL baselines, achieving up to 30% higher returns and 25% faster convergence, demonstrating the benefit of jointly modeling latent skills and adaptive temporal scope in long horizon tasks.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035274","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"CMAC-Net: a cross-modal alignment and semantic consistency optimization framework for image-text fusion in industrial defect detection","authors":"Zhuyi Rao,&nbsp;Xiya Li,&nbsp;Wei Wang","doi":"10.1007/s10489-026-07479-5","DOIUrl":"10.1007/s10489-026-07479-5","url":null,"abstract":"<div><p>Industrial defect detection requires accurate identification and localization of fine-grained anomalies, while the inconsistency between visual anomaly responses and textual defect semantics remains insufficiently addressed. To improve the semantic consistency and interpretability of industrial defect detection, this paper proposes CMAC-Net, a Cross-Modal Alignment and Consistency Network for image-text fusion. First, a dual-encoder feature extraction module models local textures, global structures, and defect-related textual semantics from industrial images and descriptions. Then, a hierarchical disentangled cross-modal alignment module separates modality-shared and modality-specific representations and establishes fine-grained correspondences between local defect regions and textual semantics. Furthermore, a text-guided semantic consistency optimization module enhances multi-level visual residual features with aligned textual information, thereby refining visual anomaly responses. Experimental results demonstrate that CMAC-Net achieves competitive performance in industrial defect detection, pixel-level localization, and image-text semantic matching. Specifically, CMAC-Net improves mIoU by 0.83% over Reverse Distillation and AUPRO by 0.92% over SimpleNet on MVTec AD, improves mIoU by 15.82% over CutPaste on VisA, and increases R@1(i2t) by 3.24% over DecAlign on Flickr30k. These results demonstrate that incorporating hierarchical cross-modal alignment and text-guided semantic refinement can effectively strengthen the correspondence between defect regions and textual descriptions, providing a more reliable framework for interpretable and semantically consistent industrial defect detection.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035083","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Complementarity-aware feature selection via weighted neighborhood probability rough sets for integrated traditional Chinese and Western medicine data in chronic renal failure prediction","authors":"Xinfang Zhang,&nbsp;Xiaoxiao Chang,&nbsp;Xiaoxia Liu,&nbsp;Bingzhen Sun,&nbsp;Xiaoli Chu","doi":"10.1007/s10489-026-07482-w","DOIUrl":"10.1007/s10489-026-07482-w","url":null,"abstract":"<div><p>Integrating Traditional Chinese Medicine (TCM) and Western medicine remains challenging because TCM data are largely unstructured and experience-based, whereas Western medicine mainly relies on quantitative clinical indicators. To address this issue, this study proposes a complementarity-aware feature selection method for integrated TCM and Western medicine data based on text mining and weighted neighborhood probability rough sets. First, a specialized TCM semantic framework is constructed based on classical literature, and SnowNLP is used to quantify textual diagnostic information. Second, a Weighted Neighborhood Probability Rough Set model (WNPRS-MSIHIS) is developed to process heterogeneous, incomplete, and multi-source medical data. In this framework, the Choquet integral and generalized Shapley values are introduced to characterize attribute importance and interactions, while granularity weights are incorporated to maintain knowledge consistency across information sources. The proposed method is validated on 3008 chronic renal failure cases. Experimental results show that the framework effectively removes redundant attributes, preserves critical diagnostic information, and improves predictive performance across multiple classifiers. The method achieves strong recall and F1-score performance, with recall reaching 0.9115 and F1-score reaching 0.8542. These results demonstrate that the proposed framework provides an effective and interpretable approach for feature selection and disease prediction in integrated TCM and Western medicine diagnosis.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035447","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Probabilistic satellite collision risk prediction via Monte Carlo ensembles","authors":"Meriem Ouari,&nbsp;Fella Manel Zerrouki,&nbsp;Seif Eddine Bouziane,&nbsp;Amel Gacem","doi":"10.1007/s10489-026-07494-6","DOIUrl":"10.1007/s10489-026-07494-6","url":null,"abstract":"<div><p>The rapid growth of resident space objects in low Earth orbit has significantly increased the risk of satellite collisions, making reliable conjunction risk forecasting a critical operational requirement. Conjunction Data Messages (CDMs), provided by the United States Space Force 18th Space Control Squadron, constitute the primary data source for collision risk assessment. However, modelling CDM data remains challenging due to severe class imbalance, a large concentration of probabilities at the numerical detection floor, and the need for both calibrated probabilistic estimates and accurate risk classification. This paper proposes a hybrid artificial intelligence framework combining Monte Carlo probabilistic modelling with ensemble machine learning. The proposed approach integrates gradient boosting models and deep neural networks within a stacked ensemble architecture, while event-level sequence models are additionally evaluated for their ability to capture temporal patterns across conjunction trajectories. A calibration strategy is introduced to improve the reliability of predicted collision probabilities. Experiments conducted on the ESA Spacecraft Collision Avoidance Challenge dataset, comprising 199,082 Conjunction Data Messages across 15,321 conjunction events, demonstrate strong predictive performance. The proposed framework achieves an F2 score of 0.808 and an AUC-ROC of 0.981 while maintaining well-calibrated probabilistic forecasts. These results demonstrate the potential of hybrid probabilistic–ensemble approaches for improving AI-driven satellite collision risk assessment.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149034447","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Semi-supervised regression via confidence-weighting and residual-connection","authors":"Liyan Liu,&nbsp;Yunfei Mao,&nbsp;Xing Lu,&nbsp;Fan Min","doi":"10.1007/s10489-026-07489-3","DOIUrl":"10.1007/s10489-026-07489-3","url":null,"abstract":"<div><p>Semi-supervised regression enhances model performance by effectively leveraging abundant unlabeled data. Popular methods enrich the training data by selecting a small number of high-confidence unlabeled data, thereby overlooking valuable information in the remaining unlabeled data. Recent methods leverage all unlabeled data by employing off-the-shelf semi-supervised regressors to generate pseudo-labels. However, they treat all pseudo-labels uniformly during training, disregarding inherent quality differences and potentially introducing significant noise. In this paper, we propose a novel semi-supervised regression neural network framework called FullReg that effectively leverages all unlabeled data while ensuring robust learning. First, we propose a confidence-weighting mechanism based on data similarity to adjust the weights of pseudo-labels in the loss. It mitigates the influence of unreliable pseudo-labels, thereby ensuring the validity of the loss and enhancing model robustness. Second, we design a residual-connection mechanism to progressively evolve model parameters across training epochs. It enhances model stability and robustness by blending parameters from previous epochs with current ones. Experiments conducted on benchmark datasets from biomedical, business, ecology, physical, and life domains, as well as a real-world solar photovoltaic dataset, show that FullReg compares favorably with eight state-of-the-art algorithms in most benchmark settings.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149035264","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}