Seunghoon Han , Hyewon Lee , Daniel Y. Lee , Sung-Soo Kim , Susik Yoon , Sungsu Lim
{"title":"Sequence-aware adaptive graph convolutional recurrent networks for traffic forecasting","authors":"Seunghoon Han , Hyewon Lee , Daniel Y. Lee , Sung-Soo Kim , Susik Yoon , Sungsu Lim","doi":"10.1016/j.knosys.2025.114533","DOIUrl":"10.1016/j.knosys.2025.114533","url":null,"abstract":"<div><div>Traffic forecasting is a crucial task for the Intelligent Transportation System (ITS). A promising research direction for improving traffic prediction is to learn dynamic graph structures incorporating the hidden dependencies from the training sequence data. However, existing works optimize these dynamic graph structures only for the training data, regarding them as static when testing with new input sequences. This constrains the forecasting model’s ability to address potential discrepancies between training and testing sequences, which may arise from unforeseen changes in the traffic environment. To address this challenge, we propose a new encoder-decoder framework for traffic forecasting, <em>S</em>equence-aware Adaptive Graph Convolutional Recurrent Networks (<span>SAGCRN</span>). The encoder augments an input sequence by exploiting spatio-temporal contexts and traffic pattern storage. Then, the decoder adaptively learns a new graph structure reflecting the augmented input sequence and uses it for prediction. To further enhance the sequence-specialized graph structure, SAGCRN optimizes the stored traffic patterns to be more discriminative. We demonstrate the superior performance of <span>SAGCRN</span> on three real-world benchmark datasets, comparing it with nine baseline models. The additional sensitivity and qualitative analyses substantiate the effectiveness of our model. For reproducibility, the source code is available at <span><span>https://github.com/gooriiie/SAGCRN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114533"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222349","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":"Toward customized model discrepancies in personalized federated learning on non-IID data","authors":"Fengrui Hao , Taihang Zhi , Tianlong Gu , Xuguang Bao","doi":"10.1016/j.knosys.2025.114522","DOIUrl":"10.1016/j.knosys.2025.114522","url":null,"abstract":"<div><div>Federated learning (FL) is a traditional framework comprising a central server and multiple local clients. In FL, a shared global model is trained for resource-constrained computing devices while preserving data privacy. However, in certain practical applications, the shared global model may exhibit poor inference performance in local clients owing to nonindependent and nonidentically distributed (non-IID) characteristics of data. To address this issue, researchers have proposed personalized FL (PFL), which involves learning a customized model for each client to mitigate the impact of weight divergences when the training datasets are non-IID. Unfortunately, existing studies fail to reveal the inherent connection between model discrepancies and non-IID data. Herein, we focus on demonstrating the relationship between weight divergences among customized models and non-IID data, and we provide a proposition to reveal the root cause of such divergences. Additionally, based on our theoretical analysis, we introduce two novel personalized FL methods, namely, PFL with neighbor clients (PFedNC) and PFL with neighbor layers (PFedNL), to address the issue of non-IID data scenarios. Theoretical convergence analysis and extensive experiments indicate that our proposed methods outperform state-of-the-art personalized algorithms in non-IID scenarios. Specifically, PFedNC achieves up to 4 % improvement in customized model accuracy, while PFedNL yields 8 %–10 % gains over multiple baselines.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114522"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222348","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}
Minrui Jiang , Yuning Yang , Xiurui Xie , Pei Ke , Guisong Liu
{"title":"Safe and effective post-fine-tuning alignment in large language models","authors":"Minrui Jiang , Yuning Yang , Xiurui Xie , Pei Ke , Guisong Liu","doi":"10.1016/j.knosys.2025.114523","DOIUrl":"10.1016/j.knosys.2025.114523","url":null,"abstract":"<div><div>Fine-tuning is critical to customizing Large Language Models (LLMs) in various applications, but it inevitably disrupts the safety alignment of the models. Current alignment methods tackle harmful fine-tuning challenges but frequently compromise model usefulness, resulting in unsatisfactory downstream task performance. To address this issue, we propose a <strong>S</strong>afe and <strong>E</strong>ffective post-fine-tuning <strong>A</strong>lignment (<strong>SEA</strong>) from a knowledge disentanglement perspective. SEA introduces a novel two-level pruning process that surgically removes harmful functionalities. We first propose a differential importance score to isolate harmful pathways at the parameter level, and then introduce a module-wise analysis to protect entangled modules, thereby robustly balancing safety and utility. Experimental results on Llama2, Gemma and Mistral demonstrate that SEA effectively mitigates safety risks while maintaining optimal fine-tuning accuracy. This work provides a practical solution to the safety-performance dilemma associated with harmful fine-tuning of LLMs.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114523"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222347","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":"The parallel visual perception network based on nonlinear spiking neural P systems for camouflaged object detection","authors":"Nan Zhou, Hong Peng, Zhicai Liu","doi":"10.1016/j.knosys.2025.114532","DOIUrl":"10.1016/j.knosys.2025.114532","url":null,"abstract":"<div><div>Numerous species have evolved camouflage through morphological adaptations that mimic environmental colors and textures, posing significant challenges for visual detection systems. Current camouflaged object detection (COD) methods remain limited in simulating biological visual mechanisms due to inadequate multi-stage cognitive modeling and weak biological correspondence in neural computations. To address these limitations, a parallel visual perception network (NSNPVPNet) based on nonlinear spiking neural P (NSNP) systems is proposed, simulating biological visual processes through three core modules: scene perception, cognitive reasoning, and decision inference module. A bio-inspired convolutional block reconstructed through NSNP systems enhances biological-computational mapping relationships. Experimental evaluations across four benchmark datasets demonstrate superior performance over twenty state-of-the-art COD methods, achieving average metric improvements of 3.2% (<span><math><msub><mi>S</mi><mi>m</mi></msub></math></span>), 2.5% (<span><math><mrow><mi>a</mi><msub><mi>E</mi><mi>m</mi></msub></mrow></math></span>), 5.4% (<span><math><msubsup><mi>F</mi><mi>β</mi><mi>w</mi></msubsup></math></span>), and 1.2% (<span><math><mi>M</mi></math></span>). These advancements validate NSNP systems’ potential in COD applications and pioneer new bio-inspired approaches for bionic visual computing. The implementation is available at: <span><span>https://github.com/Williamzhounan/NSNPVPNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114532"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222351","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}
Abudurexiti Reheman , Yingfeng Luo , Junhao Ruan , Hongyu Liu , Tong Xiao , Jingbo Zhu
{"title":"Pseudo-kNN-MT: Enhancing domain adaptability of neural machine translation via target language data","authors":"Abudurexiti Reheman , Yingfeng Luo , Junhao Ruan , Hongyu Liu , Tong Xiao , Jingbo Zhu","doi":"10.1016/j.knosys.2025.114513","DOIUrl":"10.1016/j.knosys.2025.114513","url":null,"abstract":"<div><div>Although Neural Machine Translation (NMT) has recently achieved remarkable performance improvements, it still faces challenges in domain adaptation. Previous research has focused on mitigating this issue by integrating translation knowledge from bilingual domain data. However, the limited availability of bilingual translation resources has constrained these methods in real world application. To address this inadequacy, solutions based on monolingual data, such as back-translation, have been proposed. Nevertheless, these methods often incur additional training costs due to the necessity of training reverse models to generate pseudo data. In light of this, we propose Pseudo-<span><math><mi>k</mi></math></span>NN-MT, which does not require additional training. This method creates pseudo-bilingual data pairs by retrieving semantically similar sentences from target language data and subsequently builds the <span><math><mi>k</mi></math></span>NN datastore. To effectively reduce the noise introduced by the pseudo-data, we incorporate cross-lingual retrieval distances into the <span><math><mi>k</mi></math></span>NN probability construction process. Experiments in both high-resource and low-resource machine translation scenarios across multiple domains demonstrate that our method significantly improves the domain adaptation capabilities of NMT in both settings, yielding average improvements of 6.08 and 7.70 SacreBLEU points and 0.66 and 1.62 COMET scores on the multi-domain dataset, respectively.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114513"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222171","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}
Ruihui Hou , Shencheng Chen , Yongqi Fan , Guangya Yu , Lifeng Zhu , Jing Sun , Jingping Liu , Tong Ruan
{"title":"MSDiagnosis: A benchmark and framework for evaluating large language models in multi-step clinical diagnosis","authors":"Ruihui Hou , Shencheng Chen , Yongqi Fan , Guangya Yu , Lifeng Zhu , Jing Sun , Jingping Liu , Tong Ruan","doi":"10.1016/j.knosys.2025.114524","DOIUrl":"10.1016/j.knosys.2025.114524","url":null,"abstract":"<div><div>Clinical diagnosis is critical in clinical decision-making, typically requiring a continuous and evolving process that includes primary, differential, and final diagnoses. However, most existing clinical diagnostic tasks are single-step processes, which do not align with the complex multi-step diagnostic procedures found in real clinical scenarios. In this paper, we propose MSDiagnosis, a Chinese multi-step clinical diagnostic benchmark consisting of 2225 cases from 12 departments, covering primary, differential, and final diagnosis tasks. Conventional approaches often rely on large language models (LLMs) to perform these tasks sequentially, which can lead to error propagation. To address this, we propose a two-stage diagnostic framework consisting of a forward inference module and a backward reasoning and refinement module. This framework is applied at each diagnostic stage to effectively mitigate error propagation across steps. The forward module retrieves similar cases to assist the LLM in generating an initial diagnosis. In the backward inference and refinement module, we first perform backward inference to infer the diagnostic criteria associated with the initially identified potential diseases. These criteria are then compared with the patient’s records to identify and eliminate possible misdiagnoses. Finally, the diagnostic conclusion is further refined and confirmed. Based on the MSDiagnosis, we evaluate medical LLMs (e.g., OpenBioLLM, PULSE, and Apollo2), general LLMs (e.g., DeepSeek-V3, OpenAI-O1, and GLM4), and our proposed framework. Experimental results show that our framework achieves state-of-the-art performance, demonstrating its effectiveness in multi-step diagnostic tasks. We also provide a detailed analysis and suggest future research directions for this task. Our code and data are publicly available at <span><span>https://github.com/nlper-hou/MSDiagnosis</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114524"},"PeriodicalIF":7.6,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222354","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}
Mert Sehri , Zehui Hua , Francisco de Assis Boldt , Patrick Dumond
{"title":"Selective embedding for deep learning","authors":"Mert Sehri , Zehui Hua , Francisco de Assis Boldt , Patrick Dumond","doi":"10.1016/j.knosys.2025.114535","DOIUrl":"10.1016/j.knosys.2025.114535","url":null,"abstract":"<div><div>Deep learning has revolutionized many industries by enabling models to automatically learn complex patterns from raw data, reducing dependence on manual feature engineering. However, deep learning algorithms are sensitive to input data, and performance often deteriorates under nonstationary conditions and across dissimilar domains, especially when using time-domain data. Conventional single-channel or parallel multi-source data loading strategies either limit generalization or increase computational costs. This study introduces selective embedding, a novel data loading strategy, which alternates short segments of data from multiple sources within a single input channel. Drawing inspiration from cognitive psychology, selective embedding mimics human-like information processing to reduce model overfitting, enhance generalization, and improve computational efficiency. Validation is conducted using six time-domain datasets, demonstrating that the proposed method consistently achieves high classification accuracy for many deep learning architectures while significantly reducing training times. Across multiple datasets, selective embedding consistently improves test accuracy by 20 to 30 percent compared to traditional single-channel loading strategies, while also matching or exceeding the performance of parallel multi-source loading methods. Importantly, these gains are achieved while significantly reducing training times, demonstrating both efficiency and scalability across simple and complex architectures. The approach proves particularly effective for complex systems with multiple data sources, offering a scalable and resource-efficient solution for real-world applications in healthcare, heavy machinery, marine, railway, and agriculture, where robustness and adaptability are critical.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114535"},"PeriodicalIF":7.6,"publicationDate":"2025-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145159950","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":"Monotonic learning in the PAC framework: A new perspective","authors":"Ming Li , Chenyi Zhang , Qin Li","doi":"10.1016/j.knosys.2025.114504","DOIUrl":"10.1016/j.knosys.2025.114504","url":null,"abstract":"<div><div>Monotone learning describes learning processes in which expected error consistently decreases as the amount of training data increases. However, recent studies challenge this conventional wisdom, revealing significant gaps in the understanding of generalization in machine learning. Addressing these gaps is crucial for advancing the theoretical foundations of the field. In this work, we utilize Probably Approximately Correct (PAC) learning theory to construct a theoretical error distribution that approximates a learning algorithm’s actual performance. We rigorously prove that this theoretical distribution exhibits monotonicity as sample sizes increase. We identify two scenarios under which deterministic algorithms based on Empirical Risk Minimization (ERM) are monotone: (1) the hypothesis space is finite, or (2) the hypothesis space has finite VC-dimension. Experiments on three classical learning problems validate our findings by demonstrating that the monotonicity of the algorithms’ generalization error is guaranteed, as its theoretical error upper bound monotonically converges to the minimum generalization error.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114504"},"PeriodicalIF":7.6,"publicationDate":"2025-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145159229","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":"Exploitability prediction of vulnerabilities based on heterogeneous graphs","authors":"Guo Xu, Xin Chen, Xinxin Cai, Dongjin Yu","doi":"10.1016/j.knosys.2025.114517","DOIUrl":"10.1016/j.knosys.2025.114517","url":null,"abstract":"<div><div>Vulnerability exploitability prediction is the process predicting the likelihood of being exploited in real attacks by the assessment of known software vulnerabilities. Many methods have been proposed to solve the problem of exploitability prediction. However, they generally suffer from two problems. First, they only extract features from a single vulnerability, ignoring the impact of associated vulnerabilities. Second, they usually adopt simple methods (such as concatenation) to aggregate different information, which may overlook important relationships between features. In this paper, we propose a novel exploitability prediction method based on heterogeneous graphs, called ExPreHet. First, ExPreHet defines nodes and edges to construct a heterogeneous graph. Following a series of preprocessing steps, ExPreHet generates multiple attribute vectors for each node. By implementing a restart random walk strategy, ExPreHet ensures that each node can sample all categories of neighboring nodes and group them by node category. Then, ExPreHet aggregates all the attributes of each node to generate the content vector, and each category of neighboring nodes of this node to generate a category vector. After that, the content vector and all the category vectors are aggregated to generate the final representation of the node. Finally, these final representations are input into random forest (RF) for training the classifier. To effectively assess ExPreHet, this paper conducts experiments on a dataset, which contains 66,877 vulnerabilities. The experimental results show that ExPreHet achieves 83.24 %, 83.22 %, 83.28 %, 83.25 %, and 83.24 % in terms of accuracy, precision, recall, F1-score, and area under curve (AUC), respectively. ExPreHet performs significantly better than the baseline methods.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114517"},"PeriodicalIF":7.6,"publicationDate":"2025-09-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222343","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}
Siquan Huang , Yijiang Li , Chong Chen , Leyu Shi , Wentian Cai , Ying Gao
{"title":"FedCleanse: Cleanse the backdoor attacks in federated learning system","authors":"Siquan Huang , Yijiang Li , Chong Chen , Leyu Shi , Wentian Cai , Ying Gao","doi":"10.1016/j.knosys.2025.114494","DOIUrl":"10.1016/j.knosys.2025.114494","url":null,"abstract":"<div><div>Federated learning (FL) enables multiple clients to collaboratively train an efficient deep-learning model without sharing their local data. However, due to its privacy-preserving nature, FL is vulnerable to backdoor attack, which manipulates the model behaviors on the adversary-chosen input. Existing defense methods are ineffective against sophisticated stealthy backdoors, suffering from either a low benign performance or being too specific to certain assumptions and attacks. To handle the aforementioned issues, we present FedCleanse, a novel defense mechanism to address the backdoor attack in federated learning. In this work, we study the pruning-based approach, which has been proven effective but with the need for additional data for validation and suffers from high non-IID scenarios. This paper proposes a post-aggregation approach, namely FedCleanse, to neutralize backdoor effects without needing additional clean data. Our approach identifies suspicious neurons using “neuron conductance” and subsequently suppresses them after the aggregation operation, which imposes minimal impact on benign neurons. Additionally, FedCleanse is complemented by strategic perturbations to prevent backdoor transfer. Through extensive experiments, our method demonstrates superior defense capabilities across various attack types and non-IID settings, surpassing the state-of-the-art by a large margin without compromising the main task’s performance.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"330 ","pages":"Article 114494"},"PeriodicalIF":7.6,"publicationDate":"2025-09-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222346","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}