Applied Intelligence最新文献

{"title":"Correction to: A hybrid CNN-BiLSTM model with squeeze-and-excitation attention and GSABO optimization for rock drilling time prediction in underground mines","authors":"Ding Liu, Ning Li, Qizhou Wang","doi":"10.1007/s10489-026-07436-2","DOIUrl":"10.1007/s10489-026-07436-2","url":null,"abstract":"","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148838102","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":"An interpretable video summarization model integrating self-supervised contrastive learning and reinforcement learning","authors":"Jingtao Sun,&nbsp;Jiaxing Wang","doi":"10.1007/s10489-026-07413-9","DOIUrl":"10.1007/s10489-026-07413-9","url":null,"abstract":"<div><p>Unsupervised video summarization remains challenging due to limited feature robustness, the semantic gap between low-level visual cues and high-level understanding, and insufficient transparency in annotation-free key-frame selection. To address these challenges, this paper proposes an Interpretable Self-supervised Contrastive Reinforcement Learning model for video summarization (ISCRL), which integrates self-supervised contrastive learning with reinforcement learning for annotation-free key-frame selection. ISCRL first employs a CNN backbone to extract original frame-level visual features. A SimCLR-based self-supervised contrastive learning module is then applied to these pre-extracted features to learn invariant frame-level representations with improved semantic consistency and temporal robustness. These original and invariant representations are integrated into the Deep Self-attention Recurrent Summarization Network with Reinforcement Learning (DSR-RL) framework at the state-representation level, providing stable and discriminative inputs for policy-based key-frame selection. At the reward-computation level, a dual-space semantic reward mechanism is further designed to jointly constrain diversity and consistency in both the original feature space and the invariant feature space, thereby guiding the policy toward semantically discriminative and representative key-frame selection under augmented views. For decision-level inspection, ISCRL combines self-attention with smoothed Grad-CAM + + and evaluates the resulting temporal-spatial explanations through deletion-based faithfulness, perturbation-based stability, and evidence-source informativeness analyses. Experiments on SumMe and TVSum show that ISCRL achieves competitive overall performance in terms of F1-score and rank-correlation metrics under canonical, augmented, and transfer settings, with particularly strong results on TVSum and cross-dataset transfer evaluation. Ablation, stabilization, interpretability, and complexity analyses further validate its effectiveness and practicality.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148838084","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":"Bidirectional job-shop scheduling with graph-theoretic features: a graph attention network and proximal policy optimization approach","authors":"Nanfeng Ma,&nbsp;Xifan Yao,&nbsp;Kun Hu","doi":"10.1007/s10489-026-07426-4","DOIUrl":"10.1007/s10489-026-07426-4","url":null,"abstract":"<div><p>The job-shop scheduling problem (JSSP) is a pivotal combinatorial optimization challenge with extensive industrial applications. Although deep reinforcement learning (DRL) and graph neural networks (GNN) have made notable progress in addressing JSSP, limitations in state representation, action space definition, and reward function design continue to constrain model performance and generalization capabilities. In this paper, we introduce a novel DRL-based approach that integrates bidirectional scheduling with graph-theoretic features to effectively solve JSSP. Specifically, we combine bidirectional scheduling time metrics with graph-theoretic features to form a comprehensive feature matrix, which serves as the input for reinforcement learning. This feature matrix is processed by a feature extraction network comprising multi-head self-attention and graph attention mechanisms, enabling effective capture of both temporal and structural information. The extracted features are subsequently fed into proximal policy optimization (PPO) and actor-critic networks for policy learning and state evaluation. Extensive experiments conducted on multiple public benchmark datasets (LA, TA, and Dmu) with models trained across different problem scales (10 × 10 to 30 × 20) demonstrate that our proposed method achieves superior performance compared to traditional heuristics and classic DRL methods, while maintaining competitive results against recent state-of-the-art approaches. The ablation study further validates the critical role of integrating bidirectional scheduling with graph-theoretic features, providing valuable insights for advancing DRL-based solutions to combinatorial optimization problems.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148838101","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":"Supporting and explaining stock investment decisions through Shapley Additive Explanations on LSTM and MLP-based agent models with user experience on one-decade backtesting","authors":"Iván García-Magariño,&nbsp;Javier Bravo-Agapito","doi":"10.1007/s10489-026-07415-7","DOIUrl":"10.1007/s10489-026-07415-7","url":null,"abstract":"<div><p>Investing in financial markets remains a challenging task due to their highly dynamic and uncertain nature, requiring both expertise and robust decision-support tools. This paper presents a novel framework for supporting and explaining stock investment decisions through the integration of machine learning–based agents and explainable artificial intelligence techniques. Specifically, we propose the Dashboard for Intelligent Investment with XAI (DI2XAI), an interactive simulator that incorporates Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP) models to emulate automated investment strategies, enhanced with Shapley Additive Explanations (SHAP) to provide transparent and interpretable insights into model behavior. The proposed system enables users to interact with, evaluate, and compete against AI-driven investment agents within a realistic environment based on one-decade historical backtesting, thus mitigating biases associated with short-term market conditions. In addition to explainability, DI2XAI introduces a novel analytical component that estimates the probability distribution of the time required to achieve a target return on investment under specific market conditions. A comprehensive user study involving 51 participants was conducted to assess both usability and effectiveness. Results indicate that the explainability provided by SHAP was highly valued (mean score of 5.91/7 and 5.46/7), while the proposed duration–probability estimation tool achieved an even higher evaluation (6.45/7 and 5.71/7), for respectively users with and without economical experience. Furthermore, participants improved their relative profitability compared to the market by an average of 8.71% between their initial and final simulation sessions, with Cohen’ d effect size of 0.38 and a robust statistical power of 86%. Overall, this work demonstrates how explainable AI combined with deep learning–based agents and long-term backtesting can effectively support, enhance, and rationalize stock investment decision-making in realistic settings.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837745","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":"Distributed assembly hybrid flow shop scheduling optimization based on the slime mold algorithm and reinforcement learning","authors":"Mengxin Tao,&nbsp;Xiangkun Wei,&nbsp;Xiaopeng Li,&nbsp;Yanping Zhou","doi":"10.1007/s10489-026-07419-3","DOIUrl":"10.1007/s10489-026-07419-3","url":null,"abstract":"<div><p>To address the Distributed Assembly Hybrid Flow Shop Scheduling Problem with Sequence-Dependent Setup Times (DAHFSP-SDST), this paper proposes an Improved Multi-Objective Slime Mold Algorithm (IMOSMA) to simultaneously minimize makespan and energy consumption. A hybrid initialization strategy combining chaotic mapping, Gaussian perturbation, and the Longest Processing Time (LPT) rule is adopted to enhance population diversity and initial solution quality. A dual-population co-evolution mechanism is then designed, where an elite population uses position-based crossover while a normal population learns from elite solutions to accelerate convergence. Furthermore, a Double Deep Q-Network (DDQN)-based reinforcement learning module adaptively selects among eight neighborhood operators according to the current search state and bottleneck characteristics, enabling balanced optimization of time and energy objectives. Experimental results indicate that IMOSMA consistently outperforms the comparative algorithms on the majority of test instances, highlighting its effectiveness and stable performance on the tested distributed assembly scheduling instances.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837743","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A taxonomy-driven review of surrogate-assisted evolutionary algorithms: architectural paradigms, infill strategies, optimizers, and emerging trends","authors":"Rohit Kumar,&nbsp;Susheel Kumar Joshi","doi":"10.1007/s10489-026-07422-8","DOIUrl":"10.1007/s10489-026-07422-8","url":null,"abstract":"&lt;div&gt;&lt;p&gt;Surrogate-Assisted Evolutionary Algorithms (SAEAs) have emerged as an indispensable paradigm for solving computationally expensive, high-dimensional engineering design and scientific optimization problems. However, the existing literature remains highly fragmented, frequently treating approximation models in isolation or conflating high-level macro-modeling scopes with low-level operational infill selection mechanics. To bridge this gap, this paper presents a comprehensive, taxonomy-driven synthesis of the SAEA landscape structured via a rigorous PRISMA-based literature curation workflow. A unified, multi-tiered taxonomy is introduced that explicitly decouples macro-modeling architectures, categorized into Single-Surrogate-Based Evolutionary Algorithms (SSBEAs) and Multiple-Surrogate-Based Evolutionary Algorithms (MSBEAs), from operational active-learning selection pathways. This taxonomy systematically cross-references five distinct structural subclasses, consisting of Global Surrogates (GS), Local Surrogates (LS), Hybrid Global-Local (HGL), Heterogeneous Ensembles (HE), and Multi-Regional (MRL/MRG) formats, with their core machine learning substrates, evolutionary backbones, and traditional infill criteria spanning Elitism-Based (EL), Uncertainty-Based (UB), Distance-Based (DB), and Convergence-Driven (CD) tracks. The taxonomic data analysis uncovers a sharp operational divergence between theory and practice. Specifically, the results reveal that GS loops within SSBEAs and HE frameworks within MSBEAs completely dominate the literature. While algorithmic research actively explores diverse UB, DB, and CD pathways, including theoretical MRL/MRG designs that uniquely favor space-filling DB mechanics, applied engineering deployments almost exclusively collapse onto exploitation-heavy EL selection. To analyze the behavioral trade-offs governing these SSBEA and MSBEA paradigms, a targeted, controlled empirical evaluation is incorporated under strictly bounded query budgets, formally demonstrating that optimization efficiency is a function of system-level orchestration and scheduling rather than absolute model quantity. Furthermore, a rigorous cross-paradigm synthesis of emerging trends investigates the recent infusion of next-generation artificial intelligence, including Large Language Models (LLMs), Generative Adversarial Networks (GANs), Autoencoders (AEs), and Physics-Informed Neural Networks (PINNs). This analysis uncovers a major operational shift where advanced architectures actively serve as catalysts to break the historical dependency on the traditional EL pathway, successfully introducing novel, non-traditional active-learning mechanics such as autonomous LLM-guided strategy selectors, discrete classifier-based boundary tracks, and generative exploratory loops. Concurrently, the structural mapping exposes a severe research vacuum, revealing that localized subclasses LS and MRL/MRG remain entirely unexploited within the deep learning l","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837505","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":"Landslide identification method based on enhanced DeepLabv3+ architecture: joint optimization of bottleneck attention and dual-module graph convolution","authors":"Yubin Qian,&nbsp;Zhongrong Zhang,&nbsp;Qiang Bie,&nbsp;Jinjun Li,&nbsp;Kang Lin","doi":"10.1007/s10489-026-07432-6","DOIUrl":"10.1007/s10489-026-07432-6","url":null,"abstract":"<div><p>Landslides are common and destructive geohazards, typically triggered by earthquakes and intense rainfall, and they may cause severe casualties, economic losses, and environmental damage. Rapid and accurate landslide mapping from remote sensing imagery is therefore of great importance for emergency response, post-disaster assessment, and reconstruction planning. However, landslide regions often exhibit irregular boundaries, fragmented structures, large scale variation, and strong visual confusion with surrounding backgrounds, which makes precise segmentation difficult. To address these issues, this study proposes an enhanced DeepLabv3+-based landslide semantic segmentation model, termed RSBG. The model adopts ResNet50 as the backbone, introduces attention-guided fusion for selective multi-level feature integration, uses Atrous Spatial Pyramid Pooling (ASPP) for multi-scale contextual aggregation, and combines the Bottleneck Attention Module (BAM) with the Dual Graph Convolution Network (DualGCN) for background suppression and complementary long-range dependency modeling in the spatial and channel domains. In this way, the proposed framework forms a progressive collaborative mechanism that follows a \"prioritize first, propagate next” strategy. Experimental results show that RSBG achieves 80.92% mIoU and 95.86% Accuracy on the Mengdong sub-dataset, outperforming U-Net, PSPNet, and DeepLabv3+ overall. Further ablation experiments, cross-region generalization tests, stability analysis, and efficiency evaluation demonstrate that the proposed framework performs well in terms of module collaboration, regional adaptability, and the balance between performance and efficiency. These results indicate that RSBG can effectively improve segmentation quality in complex landslide regions and shows strong potential for landslide remote sensing applications.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837659","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":"ACKGE: knowledge graph embedding with probabilistic semantic bounds for adaptive contrastive learning","authors":"Gaohao Wu,&nbsp;Xia Wang,&nbsp;Guosheng Hao,&nbsp;Xingze Lu,&nbsp;Zishu Liu","doi":"10.1007/s10489-026-07421-9","DOIUrl":"10.1007/s10489-026-07421-9","url":null,"abstract":"<div><p>Knowledge Graph Embedding (KGE) models often suffer from semantic inconsistency due to their reliance on independent triplet-level optimization, which ignores global contextual correlations. While contrastive learning has been introduced to mitigate this, existing methods typically employ rigid fixed thresholds or uniform random sampling, failing to account for the heterogeneous semantic densities across different entities. To address these limitations, we propose the Adaptive Contrastive Knowledge Graph Embedding (ACKGE) framework. The core novelty of ACKGE lies in its entity-specific moment-based modeling mechanism, which uses the empirical mean and standard deviation of entity similarities to define probabilistic semantic bounds. By deriving adaptive thresholds from these entity-specific similarity statistics, ACKGE separates high-confidence positive neighbors from semantically ambiguous “hard” negatives. This distribution-aware sampling strategy introduces additional representation-level constraints that encourage semantically related entities to form more coherent neighborhoods while maintaining discriminative boundaries. Experimental results on three benchmark datasets (FB15k-237, WN18RR, and YAGO3-10) demonstrate that ACKGE achieves competitive performance against strong baselines in link prediction tasks, particularly in handling complex relational structures.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837722","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":"ConTemPreT: Contextual temporal pre-trained transformer-based forecasting","authors":"Yunus Emre Midilli,&nbsp;Sergei Parshutin","doi":"10.1007/s10489-026-07423-7","DOIUrl":"10.1007/s10489-026-07423-7","url":null,"abstract":"<div><p>Pre-training transformer-based foundational models has recently emerged as a promising direction in the time series forecasting domain. However, most foundational models rely on a single-task pre-training approach, which limits generalization, introduces bias toward specific aspects, and reduces robustness. To address these limitations, we propose ConTemPreT, a novel transformer-based foundational model pre-trained through a sequential multi-task mechanism that combines self-supervised masked autoencoding and contrastive learning. Foundational models are pre-trained on universal datasets and subsequently fine-tuned to task-specific levels. Experimental results show that the proposed model outperforms all benchmark models across most few-shot forecasting settings and surpasses baselines on several zero-shot forecasting datasets; however, performance under the full-shot forecasting setting requires further improvement. Cross-domain ablation studies validate the generalizability of the proposed model and the individual contribution of each component.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837660","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":"Non-invasive 3D reconstruction techniques for transparent objects: A systematic review","authors":"Xinyang Cai,&nbsp;Hedan Liu,&nbsp;Yiwen Qi,&nbsp;Hao Chen","doi":"10.1007/s10489-026-07418-4","DOIUrl":"10.1007/s10489-026-07418-4","url":null,"abstract":"<div><p>Due to the complex optical properties of transparent objects, their three-dimensional (3D) reconstruction remains a critical challenge, as reflection, refraction, and multi-interface light transport invalidate traditional diffuse Lambertian assumptions. Non-invasive techniques have emerged as the primary solution for preserving object integrity, driving advancements in industrial inspection, robotic manipulation, and digital-twin construction. This review presents a PRISMA-informed systematic mapping and taxonomy of non-invasive transparent object 3D reconstruction techniques. To anchor our classification, we introduce the Optical mechanism–Cue observation–Inference paradigm–Reconstruction output (OCIR) analytical framework that links the optical mechanisms of transparent objects, the observable cues used by different methods, the corresponding inference paradigms, and the final reconstruction outputs. Based on this framework, existing approaches are organized into three main categories: Shape-from-X methods, refractive ray correspondence-based methods, and learning-based methods. Shape-from-X methods exploit indirect cues such as silhouettes, polarization, ultraviolet fluorescence, and background distortion, while refractive ray correspondence-based methods explicitly model physical light-path constraints. Learning-based methods leverage data-driven models for monocular estimation, RGB-D depth completion, and multi-view reconstruction, including multi-view stereo (MVS), neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS). We also analyze existing datasets, evaluation metrics, and the challenges involved. This review aims to provide a structured roadmap for future research on transparent object reconstruction.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 13","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148837507","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}