Expert Systems最新文献

{"title":"Small Language Models: A Systematic Review of Computational Trade-Offs, Privacy Advantages and Deployment in Intelligent Systems","authors":"Sena Dikici,&nbsp;Turgay Tugay Bilgin","doi":"10.1111/exsy.70331","DOIUrl":"https://doi.org/10.1111/exsy.70331","url":null,"abstract":"<p>This systematic review synthesizes evidence from 68 studies, including peer-reviewed journal articles, indexed conference/workshop proceedings and five remaining arXiv preprints published between 2022 and 2025, on small language models (SLMs) as computationally efficient alternatives to large language models (LLMs). Following PRISMA 2020 guidelines, we analyse optimization techniques, performance metrics, privacy–security trade-offs and deployment strategies across application domains. Parameter-efficient fine-tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) yield 2–10<span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mo>×</mo>\u0000 </mrow>\u0000 <annotation>$$ times $$</annotation>\u0000 </semantics></math> speedups while retaining 95%–99% of baseline accuracy and reducing trainable parameters by about 90%. Knowledge distillation and domain-specific fine-tuning deliver 1.5–3.5<span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mo>×</mo>\u0000 </mrow>\u0000 <annotation>$$ times $$</annotation>\u0000 </semantics></math> speedups with 85%–95% task accuracy, while quantization reduces model size by 50%–75% and achieves 2–4<span></span><math>\u0000 <semantics>\u0000 <mrow>\u0000 <mo>×</mo>\u0000 </mrow>\u0000 <annotation>$$ times $$</annotation>\u0000 </semantics></math> inference speedups with 85%–95% accuracy retention. Across benchmarks, fine-tuned SLMs attain 85%–90% accuracy on domain-specific tasks at only 10%–25% of LLM computational costs, though substantial gaps remain in complex reasoning (e.g., GPT-4: 92.13% vs. CodeT5-Large: 57.58%). Privacy analysis shows locally deployed SLMs achieve higher privacy and security scores, with healthcare, legal and IoT implementations often exceeding 9.0/10. Dataset analysis indicates that well-curated 1–50 GB domain corpora can offset smaller model capacity, with data quality outweighing volume and imbalance mitigation yielding 10%–25% gains in low-resource settings. Mobile and edge studies show LoRA plus quantization reduce latency from 500 to 50 ms, enabling 95% + efficiency gains for privacy-preserving, resource-constrained AI.</p>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 8","pages":""},"PeriodicalIF":3.4,"publicationDate":"2026-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1111/exsy.70331","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148274750","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Discover Class-Based Feature Distribution by Encoding Discrete Data for Classification","authors":"Qiang Fu,&nbsp;Yuefeng Li,&nbsp;Xiaohui Tao,&nbsp;Jianming Yong","doi":"10.1111/exsy.70288","DOIUrl":"https://doi.org/10.1111/exsy.70288","url":null,"abstract":"<p>The self-organisation map is an unsupervised learning technique that discovers patterns and relationships in data without requiring labelled training data. Inspired by the self-organisation map, Self-Organised Granular encoding has been shown to be effective for generating reliable discrete data clustering results as it is a data encoding technique that uses fuzzy sets and granularity to handle uncertain and imprecise information within discrete data. However, it is mainly useful for unsupervised learning, and its feasibility for supervised learning has not yet been studied. Also, discrete data classification is still under-researched. This paper proposes a new discrete data classification method called Transposed Fuzzy Class Granular classification. This method aims to transform discrete data into fuzzy partitions by considering all available classes and generating representations of the trained class's Transposed Fuzzy Class Granular distribution by measuring the total divergence from the average of each fuzzy class's membership degree distribution. The paper introduces a novel approach to discrete data classification by adapting class granules for classification and improving performance by tackling uncertainty, ambiguity, and the unique characteristics of discrete datasets. The study examined seven discrete datasets and compared their performance with eight commonly used classifiers as the baseline. These datasets were naturally discrete or created by discrete partitions of real datasets. The experimental results demonstrate that the proposed classifier outperforms the baseline classifiers in discrete data classification.</p>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1111/exsy.70288","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148174849","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Machine Learning-Based Loan Approval Automation: Enhancing Efficiency, Accuracy and Fairness in Credit Decision-Making","authors":"Mani Ghahremani,&nbsp;Hellen Amimo Otieno,&nbsp;Mishanil Kazreen,&nbsp;Stavros Shiaeles","doi":"10.1111/exsy.70307","DOIUrl":"https://doi.org/10.1111/exsy.70307","url":null,"abstract":"<p>Traditional loan approval processes are manual, time-consuming and susceptible to human bias. This research develops a machine learning-based system to automate loan eligibility assessment while enhancing efficiency, accuracy and fairness in credit decision-making. We developed and compared multiple supervised ML models—including Random Forest, XGBoost, stacking and voting ensembles—on a publicly available loan dataset (614 instances, 13 features). To address class imbalance (68.7% approved, 31.3% rejected), we applied the Synthetic Minority Over-sampling Technique (SMOTE). Hyperparameter tuning was performed using GridSearchCV with 5-fold cross-validation, optimising for F1-score. Model performance was evaluated using accuracy, precision, recall, F1-score and Area Under the Curve (AUC). Both Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) were applied to ensure transparency. A tuned Random Forest classifier achieved the best performance with an accuracy of 85.96%, F1-score of 87.17% and recall of 95.32%, outperforming XGBoost and ensemble methods. Cross-dataset validation on the Statlog German Credit dataset (1000 instances, 20 features) confirmed the framework's generalisability, with comparable AUC values (0.90 vs. 0.91). Overfitting analysis confirms moderate generalisation gaps, and a gender-based fairness evaluation shows all models exceed the 0.80 disparate impact threshold. This study demonstrates that a systematic framework combining rigorous hyperparameter tuning, class imbalance handling and explainable AI can create accurate, transparent and equitable loan approval systems, providing a practical blueprint for responsible AI deployment in credit decision-making.</p>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1111/exsy.70307","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148167021","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"UADC: Uncertainty-Aware Diverse Co-Training for Semi-Supervised Medical Image Segmentation","authors":"Tengfei Zhao,&nbsp;Weiyan Tong,&nbsp;Zhixiao Wang,&nbsp;Chong Fu,&nbsp;Cheng Gao","doi":"10.1111/exsy.70298","DOIUrl":"https://doi.org/10.1111/exsy.70298","url":null,"abstract":"<div>\u0000 \u0000 <p>In semi-supervised segmentation, deep co-training has emerged as a promising paradigm by utilizing prediction agreement and unlabelled-data-driven supervisory signals. Despite its effectiveness, insufficient diversity between the collaborating models often causes them to become highly consistent at an early stage, weakening the benefit of mutual learning and eventually making the framework behave similarly to self-training. In an effort to postpone this premature convergence, different initial perturbations were incorporated, but unfortunately, this resulted in a decrease in the quality of the model's pseudo-labels, which further deteriorated the model's performance. To tackle this issue, we propose an Uncertainty Aware and Diverse Co-training model (UADC) which is in line with the essential postulates of co-training. The model features a novel co-training framework working in hybrid spatial-frequency domain. Specifically, after revisiting the basic hypotheses of co-training, we devise a novel multi-view training approach that addresses the issues of dependency and role inequality inherent in traditional methods. Targeting the misconceptions in previous co-training architectures, we incorporate Segformer and frequency domain training to decouple and thereby counteract the tendency towards self-training degeneration during later training stages. Furthermore, to enhance the quality of pseudo-labels, we delve into the impact of uncertainty on model performance. We alter the structure of the model, designing a method grounded in the epistemic uncertainty of the segmentation model to optimize the quality of pseudo labels. We conduct extensive experiments on three public medical datasets of ISIC, Kvasir and CVC-ClinicDB, and compared our model with other state-of-the-art approaches, demonstrating its effectiveness and competitive performance.</p>\u0000 </div>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148174845","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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":"PAACGAN: Protocol-Aware Generative Augmentation for Robust Imbalanced Network Intrusion Detection","authors":"Mohyminul Islam,&nbsp;Ali Turab,&nbsp;Farhan Ullah","doi":"10.1111/exsy.70311","DOIUrl":"https://doi.org/10.1111/exsy.70311","url":null,"abstract":"<div>\u0000 \u0000 <p>Class imbalance remains a major challenge in network intrusion detection because rare attack classes are substantially underrepresented relative to benign traffic and frequent attack categories. This imbalance biases learning algorithms towards majority classes and reduces detection performance on critical minority attacks. To address this problem, we propose PAACGAN, a Protocol-Aware Auxiliary Classifier Generative Adversarial Network for minority-class augmentation in intrusion detection. The proposed model incorporates protocol-aware conditioning based on TCP/UDP characteristics, destination port distributions, and class-specific traffic statistics so that generated samples preserve both class identity and protocol-level consistency. The augmented training data are then used to train an XGBoost classifier for intrusion detection. Experiments on the NSL-KDD and CICIDS2017 benchmark datasets show that the proposed framework improves recall and F1-score for minority classes. On the rarest CICIDS2017 attack categories, including <i>Infiltration</i> and <i>Heartbleed</i>, the proposed method achieves an F1-score improvement of up to 12.57 percentage points. These results show that protocol-aware generative augmentation is effective for improving rare-attack detection under severe data imbalance.</p>\u0000 </div>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148167022","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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 Nuclear Power Equipment Failure Knowledge Graph Construction Method Based on the CHINESE-ROBERTa-BiLSTM-CRF Model","authors":"Tianbao Zhang,&nbsp;Kefei Zhang,&nbsp;Xun Ai,&nbsp;Sheng Liu","doi":"10.1111/exsy.70289","DOIUrl":"https://doi.org/10.1111/exsy.70289","url":null,"abstract":"<div>\u0000 \u0000 <p>This paper proposes a method for constructing a knowledge graph for nuclear power equipment failures based on the CHINESE-RoBERTa-BiLSTM-CRF model, whose goal is to automate the extraction and structured utilization of nuclear power equipment failure knowledge and provide decision support for nuclear power equipment debugging work. First, the method involves constructing annotated datasets for Named Entity Recognition (NER) and Relation Extraction (RE); Second, the CHINESE-RoBERTa-BiLSTM-CRF model is constructed. This model employs Whole Word Masking (WWM) to capture deep contextual semantics. It also utilizes a Bidirectional Long Short-Term Memory (BiLSTM) network to enhance sequence features and implements a Conditional Random Field (CRF) layer for optimized boundary recognition. Finally, the method classifies relationships between entity pairs. The resulting triples are stored in a Neo4j graph database to complete the knowledge graph construction. The nuclear power equipment fault knowledge graph constructed by this method is validated using model comparisons and case verification. Experimental results demonstrate that the NER model proposed in this paper achieves an F1 score of 94.9%, significantly outperforming the baseline model. Furthermore, practical verification confirms that the knowledge graph can accurately identify faults and provide decision support for on-site anomalies.</p>\u0000 </div>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148174847","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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":"Predicting Student Performance With an Extended Crisscross and Guided Learning Enhanced Machine Learning Model","authors":"Hongxing Gao,&nbsp;Xinsen Zhou,&nbsp;Jie Xing,&nbsp;Guoxi Liang","doi":"10.1111/exsy.70295","DOIUrl":"https://doi.org/10.1111/exsy.70295","url":null,"abstract":"<div>\u0000 \u0000 <p>Accurately predicting postgraduate student performance is crucial for effective pedagogical interventions to enhance future learning performance. The significant variability in university subjects and institutions complicates this prediction process. This paper proposes a reliable machine learning model, bEGWOA-FKNN, to address these challenges by predicting student performance based on previous academic records, thereby facilitating satisfactory graduation. The EGWOA algorithm, which introduced the extended crisscross strategy and guided learning strategy to the whale optimisation algorithm (WOA), is employed to select feature subsets from a dataset collected at Wenzhou University. Its global search capability is validated through extensive experiments, including quality analysis, ablation studies and comparisons with state-of-the-art algorithms. The bEGWOA-FKNN model, which integrates the binary version of the EGWOA (bEGWOA) algorithm with the fuzzy k-nearest neighbour, selects the most informative features and achieves superior accuracy compared to several existing feature selection methods.</p>\u0000 </div>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148174844","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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 Efficient Short Text Classification Model Based on Feature Shuffling and Attention-Enhanced With Multiview Contrastive Learning","authors":"Fangbo Liu,&nbsp;Yongqing Wu","doi":"10.1111/exsy.70312","DOIUrl":"https://doi.org/10.1111/exsy.70312","url":null,"abstract":"<div>\u0000 \u0000 <p>Short text classification remains a challenging task due to semantic sparsity, limited contextual information and the susceptibility of graph-based encodings to structural noise. Existing graph neural network and contrastive learning approaches partially mitigate these issues. They enrich contextual representations through neighbourhood aggregation and reduce structural noise through consistency constraints on representations. However, they still struggle to model heterogeneous semantic relations and often produce unstable multiview representations. To address these limitations, this paper proposes an efficient short text classification model based on Feature Shuffling and Attention-enhanced Multiview Contrastive Learning (FSAMCL). The method constructs heterogeneous graphs incorporating words, entities and part-of-speech tags, which are encoded using graph convolutional networks, while a term–document matrix models global semantic distributions. Feature-wise attention is applied to recalibrate node representations by amplifying salient semantic dimensions and suppressing irrelevant features. To enhance robustness, feature shuffling introduces lightweight, semantically coherent perturbations and singular value decomposition–based noise decomposition extracts low-rank semantic structures to generate noise-resistant augmented views. Cross-view attention then aligns and selectively integrates complementary information across heterogeneous views. Leveraging these enhanced representations, the framework jointly optimizes three contrastive objectives that reinforce invariance to perturbations, improve semantic discriminability and enhance cross-view consistency. Experiments on four benchmark datasets demonstrate the effectiveness of FSAMCL, with the largest performance gain observed on the Twitter dataset, where the model surpasses the previous best method by 3.09% in accuracy and 3.27% in F1 score. These findings highlight the contribution of FSAMCL as a robust, noise-tolerant and semantically enriched solution for short text classification.</p>\u0000 </div>","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148174846","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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":"","authors":"","doi":"","DOIUrl":"","url":null,"abstract":"","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148092149","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"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":"","authors":"","doi":"","DOIUrl":"","url":null,"abstract":"","PeriodicalId":51053,"journal":{"name":"Expert Systems","volume":"43 7","pages":""},"PeriodicalIF":2.3,"publicationDate":"2026-05-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148092148","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}