Knowledge-Based Systems最新文献

{"title":"AutoEnergy: An automated feature engineering algorithm for energy consumption forecasting with AutoML","authors":"Nasser Alkhulaifi ,&nbsp;Alexander L. Bowler ,&nbsp;Direnc Pekaslan ,&nbsp;Nicholas J. Watson ,&nbsp;Isaac Triguero","doi":"10.1016/j.knosys.2025.114300","DOIUrl":"10.1016/j.knosys.2025.114300","url":null,"abstract":"<div><div>Feature engineering (FE) plays a crucial role in Machine Learning pipelines, yet it remains a time-consuming process requiring heavy domain expertise. While Automated Machine Learning (AutoML) has automated model selection and hyperparameter tuning, it often overlooks FE, which is particularly needed in specialised domains such as Energy Consumption Forecasting (ECF). To address this limitation, we introduce AutoEnergy, a novel, domain-aware FE algorithm tailored for ECF. AutoEnergy automatically generates interpretable features from timestamps and past consumption values through rule-based transformations, integrating them with AutoML for fully automated ECF modelling while reducing human intervention. The performance of AutoEnergy was evaluated using eighteen diverse real-world energy consumption datasets spanning residential, commercial, industrial, and grid power domains. Through extensive benchmarking against baseline AutoML without FE and established FE methods, namely TSFresh (with TSEfficient and TSMinimal configurations) and FeatureTools (FT), AutoEnergy demonstrated significant improvements in both predictive accuracy and computational efficiency. AutoEnergy achieved forecasting error reductions of 19.52 % to 84.72 % compared to benchmarking methods, with strong performance on smaller datasets and statistical validation via Friedman and Wilcoxon tests. AutoEnergy demonstrated notable computational efficiency by running 1.31 and 4.41 times faster than FT and TSEff, respectively. Although 1.58 times slower than TSMin, AutoEnergy achieved 82.38 % lower forecasting errors. Integrating AutoEnergy with the state-of-the-art Tabular Prior Data Fitted Network (TabPFN) resulted in significant forecasting error reductions across test sets. These findings highlight AutoEnergy’s potential to improve AutoML performance while reducing reliance on domain expertise for FE, paving the way for fully automated ML pipelines in ECF applications.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114300"},"PeriodicalIF":7.6,"publicationDate":"2025-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144917433","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":"DNMDR: Dynamic networks and multi-view drug representations for safe medication recommendation","authors":"Guanlin Liu ,&nbsp;Xiaomei Yu ,&nbsp;Zihao Liu ,&nbsp;Shucheng Liu ,&nbsp;Xue Li ,&nbsp;Xingxu Fan ,&nbsp;Xiangwei Zheng","doi":"10.1016/j.knosys.2025.114327","DOIUrl":"10.1016/j.knosys.2025.114327","url":null,"abstract":"<div><div>Medication Recommendation (MR) is a promising research topic which booms diverse applications in healthcare and clinical domains. However, existing methods mainly rely on sequential modelling and static graphs for representation learning, which ignore the dynamic correlations in diverse medical events of a patient’s sequential visits, leading to insufficient global structural exploration on nodes. Additionally, mitigating drug-drug interactions (DDIs) is another issue determining the utility of the MR systems. To address the challenges mentioned above, this paper proposes a novel MR method with the integration of dynamic networks and multi-view drug representations (DNMDR). Specifically, weighted snapshot sequences for dynamic heterogeneous networks are constructed based on discrete visits in sequential EHRs, and all the dynamic networks are jointly trained to capture both structural correlations in diverse medical events and sequential dependency in historical health conditions, aiming to achieve comprehensive patient representations with both semantic features and structural relationships. Moreover, by combining the drug co-occurrences and adverse DDIs in the internal view of drug molecule structure and the interactive view of drug pairs, safe drug representations are available to obtain high-quality medication combination recommendations. Finally, extensive experiments on real-world datasets are conducted for performance evaluation, and the experimental results demonstrate that the proposed DNMDR method outperforms the state-of-the-art baseline models with a large margin on various metrics such as PRAUC, Jaccard, DDI rates and so on. The DNMDR model’s code is available at <span><span>https://github.com/Liuguanlin818/DNMDR</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114327"},"PeriodicalIF":7.6,"publicationDate":"2025-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144917325","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 effective two-dimensional code instance dollmaker masked convolutional network for QR code beautification","authors":"Jyoti Rathi,&nbsp;Surender Kumar Grewal","doi":"10.1016/j.knosys.2025.114329","DOIUrl":"10.1016/j.knosys.2025.114329","url":null,"abstract":"<div><div>The rapid proliferation of internet applications has made Quick Response (QR) codes indispensable in domains such as electronic ticketing, warehouse management, and online payments. However, enhancing QR code visual quality for advertising and branding purposes remains challenging due to the trade-off between aesthetic appeal and scanning reliability, as well as the inefficiency of existing beautification methods. To solve these issues, this research introduces a 2-Dimensional Code Instance Improved Dollmaker Masked Convolutional Network (2DMCN) that integrates segmentation-based region of interest extraction, an Improved Dollmaker Optimization (IDO) algorithm for visual quality enhancement, and a VGG-19-based style transfer module for customizable designs. Codeword adjustment and discrete cosine transform-based embedding are employed to maintain both data integrity and visual quality. Experimental results demonstrate that 2DMCN attains a PSNR of 55.38 dB, SSIM of 0.60, FSIM of 0.70, GMSD of 0.30, noise tolerance of 87.04%, error correction capability of 91.23%, a decoding rate of 0.87, and an average processing time of 6.2 seconds. These results confirm the proposed framework’s greater efficiency, strength, and aesthetic performance associated to prevailing approaches, making it highly suitable for practical, visually appealing, and reliable QR code applications.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114329"},"PeriodicalIF":7.6,"publicationDate":"2025-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144922760","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":"ProDG: A proxy-domain-guiding strategy for multi-source-free domain adaptation in EEG emotion recognition","authors":"Bingtao Zhou ,&nbsp;Mian Xiang ,&nbsp;Qian Ning","doi":"10.1016/j.knosys.2025.114318","DOIUrl":"10.1016/j.knosys.2025.114318","url":null,"abstract":"<div><div>Subject-independent Electroencephalogram (EEG) emotion recognition has underperformed due to significant disparities among subjects. Domain adaptation (DA) is a common solution, but traditional methods require access to target domain data, raising privacy concerns. Source-free domain adaptation offers a viable solution, however, researches on it remains unexplored. Moreover, existing methods overlooked the complementary information across source domains. To overcome this challenge, we focus on exploring the inter-domain complementarity. Our core insight is that higher-confidence predictions from source models indicate regions closer to the target domain’s distribution. Based upon, we propose <strong>Pro</strong>xy-<strong>D</strong>omain-<strong>G</strong>uiding (<strong>ProDG</strong>) strategy, which pioneers confidence-guidance to achieve privacy-preserving recognition. First, we propose a Proxy Guiding theory validating that predictions of source models with higher confidence exhibit closer distributional proximity to the target domain. Then, we propose two modules: <strong>Pr</strong>oxy <strong>M</strong>utual <strong>I</strong>nformation Alignment (<strong>PrMI</strong>) constructs a proxy domain by aggregating high-confidence predictions from source models, approximating the target-overlapping region, then each source model is aligned with proxy domain via mutual information maximization; <strong>Pr</strong>oxy <strong>P</strong>seudo-<strong>L</strong>abel Alignment (<strong>PrPL</strong>) refines clustering-based pseudo-labels using cross source confidence evaluation, enhancing supervised loss quality. The whole training process utilize only the source domain model and target data, with source data being inaccessible, ensuring privacy-preserving. Our method attains state-of-the-art accuracy on DEAP(65.3 %), SEED (85.9 %) and SEED-IV (70.4 %), surpassing privacy-preserving methods by a large margin and rivaling non-privacy-preserving approaches. ProDG validates the efficacy of confidence-based proxy guiding in multi-source-free domain adaptation. This work was conducted at the College of Electronics and Information Engineering, Sichuan University in May 2025.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114318"},"PeriodicalIF":7.6,"publicationDate":"2025-08-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144922761","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 new feature selection method using deep learning and graph representation in high-dimensional datasets","authors":"Matin Chiregi ,&nbsp;Mahdi Mazinani ,&nbsp;Mitra Mirzarezaee","doi":"10.1016/j.knosys.2025.114338","DOIUrl":"10.1016/j.knosys.2025.114338","url":null,"abstract":"<div><div>In recent decades, advances in data collection and storage have led to high-dimensional datasets containing numerous, often redundant features, which can negatively affect machine learning algorithms. Feature selection has emerged as a key solution to reduce dataset dimensionality, thereby improving computational efficiency and minimizing overfitting. Traditional feature selection models have limitations in effectively handling high-dimensional data and often overlook intricate relationships between features. Therefore, they may not fully optimize model performance and may be prone to overfitting. To address these challenges, we propose a novel feature selection method based on deep learning that can better capture complex patterns and dependencies among features in high-dimensional data. This method, which uses a deep similarity measure and graph representation, involves three phases. First, the problem is modeled as a graph using the deep similarity measure. Next, primary features are clustered through a community detection model. Finally, the most influential feature within each cluster is selected using node centrality and feature appropriateness measures. Notably, the feature selection step adopts a filter-based approach rather than relying on a learning algorithm, as is common in wrapper models. This design significantly reduces computational complexity and minimizes parameter requirements compared to previous methods. By avoiding reliance on a learning algorithm, the proposed method overcomes challenges such as high computational costs while improving accuracy. Experimental results across multiple datasets demonstrate that the proposed supervised model outperforms state-of-the-art approaches, achieving average improvements of 1.5 % in accuracy and 1.77 %, 1.87 %, and 1.81 % in precision, recall, and F1-score, respectively.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114338"},"PeriodicalIF":7.6,"publicationDate":"2025-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144917327","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":"Enhancing inductive knowledge graph completion with contextual relation topology learning","authors":"Yuxuan Lu ,&nbsp;Guojie Ma ,&nbsp;Shiyu Yang ,&nbsp;Junxiao Wang","doi":"10.1016/j.knosys.2025.114302","DOIUrl":"10.1016/j.knosys.2025.114302","url":null,"abstract":"<div><div>Knowledge graph completion (KGC) plays a crucial role in inferring missing triples within knowledge graphs (KGs), while inductive KGC extends this by enabling predictions for previously unseen entities, allowing dynamic updates in KGs. Recent methods define entity-independent features and utilize Graph Neural Networks (GNNs) to extract them from subgraphs surrounding the target triplet, which are then used to represent relational semantics and logical rules for reasoning. However, the inductive capabilities of existing work is limited as they consider limited entity-independent features. To address this issue, we introduce a novel <strong>C</strong>ontextual <strong>R</strong>elation <strong>T</strong>opology <strong>L</strong>earning-based GNN framework for inductive KGC, namely <strong>CRTL</strong>, which considers a broader range of entity-independent features. We observe that <strong>subgraph structural features, relation correlation patterns</strong>, and <strong>entity-relation interactions</strong> are crucial entity-independent features for inductive KGC. Moreover, relation correlation patterns and entity-relation interactions are complementary. Specifically, we construct enclosing subgraphs to extract subgraph structural features, relational graphs to model the correlations between relations, and context subgraphs to capture the interactions between entities and relations. In addition, we design a scoring function that dynamically adjusts the contributions of these features. Our extensive experiments on benchmark datasets reveal that CRTL surpasses current state-of-the-art methods, demonstrating improvements of 9.68 % on WN18RR v1 and 12.86 % on FB15K-237 v1 when compared to the suboptimal results.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114302"},"PeriodicalIF":7.6,"publicationDate":"2025-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144933461","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":"Knowledge distillation with predicted depth for robust and lightweight face presentation attack detection","authors":"Muhammad Shahid Jabbar ,&nbsp;Taha Hasan Masood Siddique ,&nbsp;Kejie Huang ,&nbsp;Shujaat Khan","doi":"10.1016/j.knosys.2025.114325","DOIUrl":"10.1016/j.knosys.2025.114325","url":null,"abstract":"<div><div>Face Presentation Attack Detection (FacePAD) is critical for safeguarding face recognition systems against spoofing attempts, including printed photos, video replays, and 3D masks. However, many existing approaches struggle with generalization across diverse attack types and real-world conditions. In this study, we propose a dual-branch deep learning framework that leverages both RGB images and synthetically predicted depth maps to improve anti-spoofing robustness and accuracy. A monocular depth estimation network is used to generate depth cues from a single RGB image, which are then processed in parallel with the original image through two distinct branches of a convolutional neural network. The extracted features-texture-based from RGB and structure-aware from depth-are fused via concatenation to facilitate more discriminative spoof detection. Extensive experiments on four benchmark datasets demonstrate that our method achieves state-of-the-art performance, reducing HTER to 0 % on Replay-Attack and Replay-Mobile, and 1.023 % on ROSE-Youtu. Similarly, an ACER of 0.56 % is achieved on OULU-NPU, while maintaining computational efficiency. Furthermore, we introduce a knowledge distillation scheme to compress the dual-branch model into a lightweight single-branch variant suitable for real-time deployment in mobile authentication, surveillance, and biometric access control scenarios.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114325"},"PeriodicalIF":7.6,"publicationDate":"2025-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144933463","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":"HSAE: Hierarchical structure augment embedding for various knowledge graph completion","authors":"Yifan Xue ,&nbsp;Wanqiang Cai ,&nbsp;Yingyao Ma ,&nbsp;Lotfi Senhadji ,&nbsp;Huazhong Shu ,&nbsp;Jiasong Wu","doi":"10.1016/j.knosys.2025.114320","DOIUrl":"10.1016/j.knosys.2025.114320","url":null,"abstract":"<div><div>Knowledge Graph Completion (KGC) addresses the task of reasoning over existing facts to predict missing relationships, serving as a fundamental component for downstream applications including question answering systems and personalized recommendation engines. Over the years, the KGC field has evolved into specialized tasks, including static KGC, temporal KGC, hyper KGC, and few-shot KGC, each requiring specialized methodologies. Although previous methods have utilized Generative Language Models (GLMs) to theoretically support multi task compatibility, their performance remains suboptimal compared to task-specific models. This limitation stems from their inability to effectively integrate structural and textual information, leading to a fine-grained structure-text gap. To address this challenge, we propose HSAE, a novel two-stage framework that hierarchically aligns structural and textual modalities, first at the coarse-grained entity level and then at the fine-grained token level. In the first stage, Entity-Level Structure Augment, we transform structural embeddings into tree-shaped entity classifications, enriching entity representations with explicit structural information. This augmentation provides global structural guidance during beam search, ensuring that generated sequences adhere to the underlying knowledge graph topology. In the second stage, Token-Level Structure Augment, we introduce a cross-modal alignment module that dynamically fuses structural embeddings with token-level predictions. By aligning structural and textual representations at the token level, HSAE ensures that each decoding step is informed by both structural and textual coherence. Experiments on eight benchmarks demonstrate that HSAE outperforms competitive baselines across multiple KGC tasks. The data and code are released at <span><span>https://anonymous.4open.science/r/HSAE-main/README.md</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114320"},"PeriodicalIF":7.6,"publicationDate":"2025-08-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144917324","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":"DFUSegNet: Boundary-aware hierarchical attentive fusion network with adaptive preprocessing for diabetic foot ulcer segmentation","authors":"Tushar Talukder Showrav ,&nbsp;Muhammad Zubair Hasan ,&nbsp;Md Kamrul Hasan","doi":"10.1016/j.knosys.2025.114323","DOIUrl":"10.1016/j.knosys.2025.114323","url":null,"abstract":"<div><div>Diabetic Foot Ulcers (DFUs) are a severe complication of diabetes, often leading to lower limb amputation and increased patient morbidity. Accurate segmentation of DFUs is essential for effective wound assessment, treatment planning, and healing monitoring. This paper introduces a novel deep learning framework, DFUSegNet, for accurate segmentation of DFUs and other chronic wounds. The proposed architecture seamlessly integrates a learnable image preprocessor (LIP) to enhance input quality and a hierarchical encoder for capturing multiscale and multiresolution wound features. A boundary enhancer (BE) sharpens ulcer edges, while the multiresolution positional attention (MPA) module emphasizes critical spatial details. Extracted features by the encoder are refined through a local-global feature aggregation (LGFA) module before being processed by a dual-mode attention-guided hierarchical decoder, ensuring precise and robust segmentation. Extensive quantitative and qualitative evaluations on the DFUC, FUSeg, and AZH Wound datasets showcase the superior performance of DFUSegNet, achieving state-of-the-art IoU/F1-scores (in %) of 60.06/70.78 on DFUC, 79.06/85.76 on FUSeg, and 81.21/87.28 on AZH. Interpretability analysis further highlights the effectiveness of our MPA, BE modules, and dual-mode attention-guided decoder in progressively extracting intricate ulcer features. Despite encountering some anomalies in the datasets, DFUSegNet demonstrates immense potential for integration into knowledge-based systems within clinical workflows and telemedicine, enabling automated, high-precision DFU segmentation to support early diagnosis and effective wound management. While promising results validate its effectiveness, successful clinical deployment will require large, accurately annotated DFU datasets, laying the foundation for future advancements in automated DFU segmentation. Source code: <span><span>https://github.com/tushartalukder/DFUSegNet</span><svg><path></path></svg></span></div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114323"},"PeriodicalIF":7.6,"publicationDate":"2025-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144917321","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":"Question answering over spatio-temporal knowledge graph","authors":"Xinbang Dai,&nbsp;Huiying Li,&nbsp;Nan Hu,&nbsp;Yongrui Chen,&nbsp;Rihui Jin,&nbsp;Huikang Hu,&nbsp;Guilin Qi","doi":"10.1016/j.knosys.2025.114314","DOIUrl":"10.1016/j.knosys.2025.114314","url":null,"abstract":"<div><div>Spatio-temporal knowledge graphs (STKGs) enhance traditional KGs by integrating temporal and spatial annotations, enabling precise reasoning over questions with spatio-temporal dependencies. Despite their potential, research on spatio-temporal knowledge graph question answering (STKGQA) remains limited. This is primarily due to the lack of datasets that simultaneously contain spatio-temporal information, as well as methods capable of handling implicit spatio-temporal reasoning. To bridge this gap, we introduce the spatio-temporal question answering dataset (STQAD), the first comprehensive benchmark comprising 10,000 natural language questions that require both temporal and spatial reasoning. STQAD is constructed with real-world facts containing spatio-temporal information, ensuring that the dataset reflects practical scenarios. Furthermore, our experiments reveal that existing KGQA methods underperform on STQAD, primarily due to their inability to model spatio-temporal interactions. To address this, we propose the spatio-temporal complex question answering (STCQA) method, which jointly embeds temporal and spatial features into KG representations and dynamically filters answers through constraint-aware reasoning. STCQA achieves state-of-the-art performance, significantly outperforming existing baselines. Our work not only provides a valuable resource for future research but also advances the field by offering a robust baseline for answering complex spatio-temporal questions.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"329 ","pages":"Article 114314"},"PeriodicalIF":7.6,"publicationDate":"2025-08-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144902497","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}