Knowledge-Based Systems最新文献

{"title":"A topology-aware multiscale feature fusion network for EEG-based motor imagery decoding","authors":"Chaowen Shen,&nbsp;Akio Namiki","doi":"10.1016/j.knosys.2025.114540","DOIUrl":"10.1016/j.knosys.2025.114540","url":null,"abstract":"<div><div>Motor imagery electroencephalography (MI-EEG) decoding is a crucial component of brain-computer interface (BCI) systems, serving as a valuable tool for motor function rehabilitation and fundamental neuroscience research. However, the strong nonlinearity and non-stationarity of MI-EEG signals make achieving high-precision decoding a challenging task. Current deep learning methods primarily extract the spatiotemporal features of MI-EEG signals while neglecting their potential association with spectral-topological features, thereby limiting the ability to integrate multidimensional information. To address these limitations, this paper proposes a Topology-Aware Multiscale Feature Fusion network (TA-MFF network) for MI-EEG signal decoding. Specifically, we designed a Spectral-Topological Data Analysis-Processing (S-TDA-P) module that leverages persistent homology features to analyze the spatial topological relationships between EEG electrodes and the persistent patterns of neural activity. Then, the Inter Spectral Recursive Attention (ISRA) mechanism is employed to model the correlations between different frequency bands, enhancing critical spectral features while suppressing irrelevant noise. Finally, the Spectral-Topological and Spatio-Temporal Feature Fusion (SS-FF) Unit is employed to progressively integrate topological, spectral, and spatiotemporal features, capturing dependencies across different domains. The experimental results show that the classification accuracy of the proposed model in BCIC-IV-2a, BCIC-IV-2b, and BCIC-III-Iva is 85.87 %, 90.2 %, and 80.5 %, respectively, outperforming the most advanced methods.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114540"},"PeriodicalIF":7.6,"publicationDate":"2025-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222173","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":"Few-Shot hyperspectral image classification with mamba and manifold convolution fusion network","authors":"Heling Cao ,&nbsp;Yanlong Guo ,&nbsp;Yonghe Chu ,&nbsp;Yun Wang ,&nbsp;Junyi Duan ,&nbsp;Peng Li","doi":"10.1016/j.knosys.2025.114531","DOIUrl":"10.1016/j.knosys.2025.114531","url":null,"abstract":"<div><div>Efficient modeling of global-local features is crucial for hyperspectral image (HSI) classification. The mamba network demonstrates strong capability in capturing global dependencies in HSI classification tasks, primarily utilizing a state-space model to extract first-order statistical features of spectral-spatial information in euclidean space, providing an initial representation of data characteristics. However, under few-shot conditions, fully exploiting effective features from limited samples and overcoming challenges such as class overlap and feature space sparsity caused by the insufficient extraction of second-order statistical features in riemannian space remain major research challenges. Therefore, we propose a dual branch manifold convolution-mamba network (DBMCMamba) for HSI classification. Specifically, it adaptively fuses forward and backward information through the vision mamba (Vim) block and utilizes the S6 module to extract global information, thereby enhancing global feature extraction capability. Meanwhile, the manifold convolution module extracts first-order statistical features of spectral-spatial information through convolutional layers and learns second-order statistics via the SPD manifold to strengthen DBMCMamba’s local feature representation under few-shot conditions. Finally, global and local features are fused for classification, effectively improving the accuracy and performance of HSI classification. On the Indian Pines, Pavia University, HongHu, and HanChuan datasets, DBMCMamba achieved classification accuracies of 95.23 %, 95.80 %, 95.58 %, and 94.93 %, respectively. Experimental results show that DBMCMamba demonstrates significant performance improvements compared to the state-of-the-art classification models. The code will be available online at <span><span>https://github.com/ASDFFGG121EAA/DBMCMamba</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114531"},"PeriodicalIF":7.6,"publicationDate":"2025-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222170","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":"Heterogeneous graph collaborative representation learning for drug-related microbe prediction with attentive fusion and reciprocal distillation","authors":"Yanbu Guo ,&nbsp;Quanming Guo ,&nbsp;Shengli Song ,&nbsp;Yihan Wang ,&nbsp;Jinde Cao","doi":"10.1016/j.knosys.2025.114548","DOIUrl":"10.1016/j.knosys.2025.114548","url":null,"abstract":"<div><div>Microbes are microorganisms with biological molecules and have significant therapeutic potential for treating diseases, underscoring the need for computational methods to screen microbes targeting disease-associated drugs. However, the computational methods often consider node embedding or structure features between microbes and drugs, and have a severe class imbalance problem inherent in sparse association data. In this work, we proposed a heterogeneous graph collaborative representation learning model that combines the merits of attentive fusion and reciprocal distillation for drug-related microbe prediction. First, we constructed the heterogeneous biological information and meta-path-induced graphs of microbes and drugs. Then, a topological structure feature encoder is devised to extract complex topological and semantic interaction patterns from heterogeneous biological graphs with microbes and drugs, while an efficient transformer concurrently extracts discriminative semantic and structural information based on the graph position information of nodes. Next, a reciprocal distillation schema is developed to mitigate the adverse effects of the data imbalance problem, and enable the distribution consistency of the model between topological and semantic information extraction. Moreover, we devised a dual collaborative feature fusion schema that combines graph topological and dual meta-path-based semantic features to obtain the discriminative features of microbes and drugs. Through reciprocal distillation, an efficient optimization function focuses on hard-to-classify samples of drug-related microbes via discriminative features. Extensive experiments demonstrate that our model could deal with the association sparsity problem and extract more semantics and structure. Meanwhile, case studies indicate that our model could discover reliable candidate microbes associated with a special drug.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114548"},"PeriodicalIF":7.6,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222223","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":"Dynamic pick-up point recommendation with multi-modal deep forest and incentive-based adaptive Kuhn-Munkres Algorithm","authors":"Yuhan Guo ,&nbsp;Rushi Zhu ,&nbsp;Wenhua Li ,&nbsp;Youssef Boulaksil ,&nbsp;Hamid Allaoui","doi":"10.1016/j.knosys.2025.114543","DOIUrl":"10.1016/j.knosys.2025.114543","url":null,"abstract":"<div><div>Recommendations for optimal pick-up points significantly enhance service efficiency, reduce economic and temporal costs, and alleviate traffic congestion. However, spatiotemporal imbalance between ride-hailing supply and passenger demand presents significant challenges. Current models often overlook critical influencing factors such as passenger satisfaction, travel environment, and travel cost factors. Moreover, solution algorithms, including exact algorithms and heuristics, struggle to achieve global optimality and computational efficiency in large-scale scenarios. This study introduces a comprehensive mathematical model that incorporates four key influencing factors: passenger walking distance, passenger waiting time, traffic conditions, and estimated ride-hailing fare. The solution approach consists of a novel pick-up point evaluation algorithm and an incentive-based adaptive Kuhn-Munkres matching algorithm. The evaluation algorithm employs a multi-modal decision tree structure, enhanced by deep learning techniques to improve the accuracy of pick-up point evaluations. The matching algorithm features a multi-scenario adaptive mechanism that dynamically adjusts edge weights and selects optimal edges for augmentation under various conditions and strategies, thereby ensuring globally optimal matching of passengers and pick-up points. Extensive experiments on large-scale real-world datasets validate the superior performance of the evaluation and matching algorithms, especially in handling large-scale instances. The developed model and algorithms assist ride-hailing platforms in optimizing operations, enhancing service quality, increasing profitability, and improving cost management.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114543"},"PeriodicalIF":7.6,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222221","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":"Uncovering novel scientific insights with a synergistic GNN-LLM framework","authors":"Qingqing Wang ,&nbsp;Derui Lyu ,&nbsp;Qiuju Chen","doi":"10.1016/j.knosys.2025.114527","DOIUrl":"10.1016/j.knosys.2025.114527","url":null,"abstract":"<div><div>The exponential growth of scientific literature demands intelligent systems capable of uncovering emerging knowledge associations and fostering creativity. While graph neural networks (GNNs) excel at modeling literature structures, their static temporal modeling and lack of semantic awareness limit the discovery of interpretable signals. Conversely, large language models (LLMs) offer deep semantic reasoning but struggle to find non-obvious, structurally-grounded patterns without structured input. To address these limitations, this paper proposes a multi-stage GNN-LLM framework that integrates structural pattern recognition and semantic interpretation for scientific knowledge discovery. The framework begins with a Semantic-Enhanced Temporal Graph Network (SE-TGN), which embeds paper-level semantic information into an event-based temporal GNN to identify emerging keyword associations. These structurally grounded candidates are refined through the Contextual Re-ranking and Evaluation Framework (CREF), which leverages LLM capabilities to assess contextual novelty and relevance. Finally, the Generative Interpretation and Contextualization (GIC) produces human-readable explanations and research prompts to support innovation. Experiments in two scientific domains demonstrate the effectiveness of the framework in discovering semantically rich, contextually grounded, and forward-looking knowledge associations, illustrating its potential to support interpretable and creativity-driven scientific exploration.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114527"},"PeriodicalIF":7.6,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222226","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 decision-based heterogenous graph attention network for multi-class fake news detection","authors":"Batool Lakzaei,&nbsp;Mostafa Haghir Chehreghani,&nbsp;Alireza Bagheri","doi":"10.1016/j.knosys.2025.114499","DOIUrl":"10.1016/j.knosys.2025.114499","url":null,"abstract":"<div><div>A promising tool for addressing fake news detection is Graph Neural Networks (GNNs). However, most existing GNN-based methods rely on binary classification, categorizing news as either real or fake. Additionally, traditional GNN models use a static neighborhood for each node, making them susceptible to issues like over-squashing. In this paper, we introduce a novel model named Decision-based Heterogeneous Graph Attention Network (DHGAT) for fake news detection in a semi-supervised setting. DHGAT effectively addresses the limitations of traditional GNNs by dynamically optimizing and selecting the neighborhood type for each node in every layer. It represents news data as a heterogeneous graph where nodes (news items) are connected by various types of edges. The architecture of DHGAT consists of a decision network that determines the optimal neighborhood type and a representation network that updates node embeddings based on this selection. As a result, each node learns an optimal and task-specific computational graph, enhancing both the accuracy and efficiency of the fake news detection process. We evaluate DHGAT on the LIAR dataset, a large and challenging dataset for multi-class fake news detection, which includes news items categorized into six classes. Our results demonstrate that DHGAT outperforms existing methods, improving accuracy by approximately 4% and showing robustness with limited labeled data.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114499"},"PeriodicalIF":7.6,"publicationDate":"2025-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222352","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":"NPMFF-Net: A training-free unified framework for point cloud classification and segmentation","authors":"Hualong Zeng ,&nbsp;Haijiang Zhu ,&nbsp;Huaiyuan Yu ,&nbsp;Mengting Liu ,&nbsp;Ning An","doi":"10.1016/j.knosys.2025.114529","DOIUrl":"10.1016/j.knosys.2025.114529","url":null,"abstract":"<div><div>Non-parametric networks have shown promise for understanding point clouds due to their training-free nature and low computational cost. However, existing methods such as Point-NN and Seg-NN underutilize geometric and frequency information. Although these methods demonstrate superior accuracy, we found that the potential features of point clouds can still be explored in depth. In this work, we revisit non-parametric networks and propose the Non-Parametric Multi-scale Feature Fusion Network (NPMFF-Net), a model designed to unify spatial and frequency information in point cloud analysis, featuring training-free components. The key is Plücker coordinates Encoding and Fourier Feature Mapping, combining geometric information with high-frequency features. We propose a non-parametric attention module to integrate contextual information and k-adaptive normal pooling to aggregate multi-scale features. Extensive experiments on the ModelNet10/40, ScanObjectNN, ShapeNetPart, S3DIS, and ScanNet datasets demonstrate the superiority of NPMFF-Net in point classification and segmentation tasks. We surpass Point-NN by 8.2 % OA and Seg-NN by 5.8 % OA on ModelNet40 for classification, while also achieving a 2.7 % improvement in mean IoU over Point-NN on ShapeNetPart for part segmentation.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114529"},"PeriodicalIF":7.6,"publicationDate":"2025-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222227","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":"Node importance estimation leveraging LLMs for semantic augmentation in knowledge graphs","authors":"Xinyu Lin ,&nbsp;Tianyu Zhang ,&nbsp;Chengbin Hou ,&nbsp;Jinbao Wang ,&nbsp;Jianye Xue ,&nbsp;Hairong Lv","doi":"10.1016/j.knosys.2025.114521","DOIUrl":"10.1016/j.knosys.2025.114521","url":null,"abstract":"<div><div>Node Importance Estimation (NIE) is a task that quantifies the importance of nodes in a graph. Recent research has investigated to exploit various information from Knowledge Graphs (KGs) to estimate node importance scores. However, the semantic information in KGs could be insufficient, missing, and inaccurate, which would limit the performance of existing NIE models. To address these issues, we leverage Large Language Models (LLMs) for semantic augmentation thanks to the LLMs’ extra knowledge and ability of integrating knowledge from both LLMs and KGs. To this end, we propose the LLMs Empowered Node Importance Estimation (LENIE) method to enhance the semantic information in KGs for better supporting NIE tasks. To our best knowledge, this is the first work incorporating LLMs into NIE. Specifically, LENIE employs a novel clustering-based triplet sampling strategy to extract diverse knowledge of a node sampled from the given KG. After that, LENIE adopts the node-specific adaptive prompts to integrate the sampled triplets and the original node descriptions, which are then fed into LLMs for generating richer and more precise augmented node descriptions. These augmented descriptions finally initialize node embeddings for boosting the downstream NIE model performance. Extensive experiments demonstrate LENIE’s effectiveness in addressing semantic deficiencies in KGs, enabling more informative semantic augmentation and enhancing existing NIE models to achieve the state-of-the-art performance. The source code of LENIE is freely available at <span><span>https://github.com/XinyuLin-FZ/LENIE</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114521"},"PeriodicalIF":7.6,"publicationDate":"2025-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222172","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 multi-modal document distillation with energy-weighted supervision","authors":"Jen-Chun Chang ,&nbsp;Chia-Cheng Lee ,&nbsp;Chung-Fu Lu ,&nbsp;Victor R.L. Shen","doi":"10.1016/j.knosys.2025.114542","DOIUrl":"10.1016/j.knosys.2025.114542","url":null,"abstract":"<div><div>As large multi-modal document models (e.g. LayoutLMv3) grow increasingly complex, knowledge distillation (KD) has become essential for practical deployment. EnergyKD enhances conventional logit-based KD by adjusting temperature per sample using energy scores. However, it still misleads students when teacher predictions are incorrect on high energy (i.e. low confidence) inputs. Although High-Energy Data Augmentation (HE- DA) is introduced to address this issue, it adds significant training overhead. In this work, we propose Energy-Weighted Supervision (EWS), a general-purpose supervision augmentation framework that builds upon an energy-based sample stratification mechanism. EWS dynamically adjusts the balance between hard-label and soft-label losses according to each sample’s energy score, thereby increasing the likelihood that the student model receives accurate and corrective supervision, without requiring additional data augmentation or training overhead. Our experiments demonstrate that EWS effectively improves the performance of various KD methods. On the harder FUNSD benchmark, EWS yields the largest gains (+2.35 F1), while on CORD and SROIE the improvements are smaller but consistently positive (up to +0.84 and +0.11 F1, respectively), confirming broad applicability across KD paradigms. Especially, when applied to EnergyKD, EWS addresses its core limitation, namely, the misleading influence of sharpened teacher outputs on high energy samples, by allocating greater weight to hard-label signals. Conversely, for low energy samples, EWS preserves soft-label emphasis to fully exploit the teacher’s informative predictions. Compared to conventional logit-based KD, EnergyKD, and even HE-DA, our energy-guided loss modulation approach consistently improves student performance across multiple documents understanding benchmarks, without additional training cost. To the best of our knowledge, this is the first framework in multi-modal document distillation that simultaneously integrates energy-aware temperature scaling and dynamic supervision weighting, offering a promising direction for future research and deployment on resource-limited devices.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114542"},"PeriodicalIF":7.6,"publicationDate":"2025-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222345","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":"Towards robust infrared small target detection: A feature-enhanced and sensitivity-tunable framework","authors":"Jinmiao Zhao ,&nbsp;Zelin Shi ,&nbsp;Chuang Yu ,&nbsp;Yunpeng Liu ,&nbsp;Yimain Dai","doi":"10.1016/j.knosys.2025.114519","DOIUrl":"10.1016/j.knosys.2025.114519","url":null,"abstract":"<div><div>Recently, single-frame infrared small target (SIRST) detection technology has attracted widespread attention. Different from most existing deep learning-based methods that focus on improving network architectures, we propose a feature-enhanced and sensitivity-tunable (FEST) framework, which is compatible with existing SIRST detection networks and further enhances their detection performance. The FEST framework improves the model’s robustness from two aspects: feature enhancement and target confidence regulation. For feature enhancement, we employ a multi-scale fusion strategy to improve the model’s perception to multi-scale features of multi-size targets, and design an edge enhancement difficulty mining (EEDM) loss to guide the network to continuously focus on challenging target regions and edge features during training. For target confidence regulation, an adjustable sensitivity (AS) strategy is proposed for network post-processing. This strategy enhances the model’s adaptability in complex scenarios and significantly improves the detection rate of infrared small targets while maintaining segmentation accuracy. Extensive experimental results show that our FEST framework can effectively enhance the performance of existing SIRST detection networks. The code is available at <span><span>https://github.com/YuChuang1205/FEST-Framework</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114519"},"PeriodicalIF":7.6,"publicationDate":"2025-09-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222174","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}