Expert Systems最新文献

This work highlights the critical need for highly reliable systems across various fields of science and technology. It emphasises the significance of achieving maximum reliability while operating within constraints such as cost, weight and volume. To address this challenge, the study introduces an innovative approach to enhance the reliability of Cold Standby systems. The proposed method involves incorporating a combination of cold standby components (RRAP-CM-CS) with advanced optimisation techniques, specifically utilising a hybrid of particle swarm optimisation and grey wolf optimiser (HPSGWO). The results obtained from simulations and real-world tests demonstrate a substantial improvement in the reliability of benchmark systems. The approach not only enhances system reliability but also surpasses the performance of traditional methods. This paper provides valuable insights into a practical and effective strategy for strengthening systems by intelligently mixing components and leveraging optimisation strategies.

Understanding sarcasm in online communication is crucial for accurately gauging public opinion. Since sarcasm conveys the opposite of its literal meaning, it throws a wrench into automated sentiment analysis tools. This has fuelled the development of new methods to identify sarcasm in text and speech. Researchers are turning to advanced techniques like word embedding (Word2Vec), Transformers (BERT) and deep learning models (RNNs) to improve the accuracy of sarcasm detection systems. Sarcasm prevalence in social media poses a challenge to sentiment analysis accuracy. This review investigates recent advancements in sarcasm detection using machine learning and deep learning techniques. We analyse studies published between 2018 and 2023, identified through a comprehensive search of Google Scholar, ScienceDirect, SpringerLink, ResearchGate and Semantic Scholar. Following PRISMA guidelines, 37 studies were selected based on inclusion/exclusion criteria. Our analysis explores four research questions and summarises the utilised datasets, algorithms and the significance of multi-modality and context in sarcasm detection. The review concludes that deep learning approaches achieve superior performance compared with other methods. In conclusion, we found that making use of multimodality significantly enhances the performance of models in sarcasm detection. The multi-modality model using the deep learning technique Bidirectional Gated Recurrent Unit (BiGRU) achieved the highest performance with an accuracy of 99.1%. Furthermore, it highlights the potential of multi-modal integration for enhanced accuracy in sarcasm detection.

This study proposes a theory-informed learning analytics approach for predicting at-risk students using large-scale behavioural, demographic and academic data. Building on engagement theory and self-regulated learning, we engineer pedagogically grounded behavioural indicators that move beyond raw click counts. These indicators include multidimensional engagement measures, temporal regularity and an antiprocrastination score derived from assessment submission patterns. Using the Open University Learning Analytics Dataset (OULAD), comprising 32,593 students across 22 courses, we reformulate the prediction task as a binary classification problem (Pass vs. At-Risk) and compare three machine learning algorithms: Support Vector Machine (SVM), Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Models are evaluated at four quarter-based checkpoints over the semester to investigate temporal dynamics and opportunities for timely intervention. Results show that XGBoost consistently outperforms RF and SVM in accuracy, recall, precision and ROC AUC, while behavioural features overwhelmingly dominate demographic and academic variables in predictive importance. Temporal analysis reveals that model performance improves substantially from the first to the third quarter, with mid-semester predictions offering the best trade-off between accuracy and time remaining for effective support. The findings demonstrate the value of theory-driven feature engineering and temporally sensitive evaluation in designing early-warning systems that are both accurate and pedagogically actionable.

Credit Risk Prediction (CRP) is crucial for financial stability but faces challenges from high-dimensional, dynamic data streams, which can cause redundancy and obscure temporal risk patterns. To address these issues, this study proposes an integrated framework that combines adaptive time-aware feature augmentation with an Improved Binary Harris Hawks Optimization (IBHHO) algorithm for feature selection. Firstly, the original credit features are augmented using four classical statistical approaches, and a Bayesian optimization mechanism is also employed to select parameters for temporal variables. Then, the IBHHO algorithm, which was developed based on Tent map initialization, nonlinear energy update, and a Whale Optimization Algorithm (WOA) fusion strategy, is introduced to identify an optimal feature subset. The integrated framework has been tested on a real Chinese dataset and two public datasets. On the dataset of Chinese small and medium-sized enterprises (SMEs), the framework achieved a prediction accuracy of 0.961 and an AUC value of 0.882, while reducing the characteristic dimension by 77.3%. These results show that the proposed framework provides a potential solution to the problem of dynamic CRP, especially for financial institutions that manage complex credit information.

Precipitation nowcasting plays a critical role in agriculture, transportation, urban planning, and emergency response. With the growing emphasis on sustainable climate resilience, AI-driven approaches have become vital for enhancing the accuracy and timeliness of meteorological forecasting. However, most existing deep learning models rely on a single data source, employ fixed-scale feature extraction, and lack effective spatio-temporal attention mechanisms. To address these issues, we propose a multi-source data fusion Transformer, termed FusionFormer, for precipitation nowcasting. The FusionFormer integrates a Multi-Source Fused Embedding (MSFE) module to effectively capture both fine-grained local features and large-scale patterns. It utilises parallel multi-size convolutional kernels and multi-resolution branches. Additionally, a Multi-scale Spatio-Temporal Attention (MSTA) module dynamically identifies the movement trajectory and intensity variations of precipitation systems. We also constructed a high-resolution, spatiotemporally aligned multi-source meteorological dataset for Shandong Province. Experiments on this dataset demonstrate that FusionFormer outperforms state-of-the-art methods in both objective and subjective evaluations.

With the exponential growth of online reviews on e-commerce platforms, efficiently identifying helpful reviews has become increasingly critical for supporting consumer decision-making and mitigating information overload. The task of Review Helpfulness Prediction (RHP) aims to address this challenge by automatically filtering high-quality and reliable content from massive volumes of user-generated reviews. While earlier studies have explored this task through both feature-based machine learning and deep learning models, these approaches often struggle to capture the complex linguistic nuances and contextual dependencies inherent in review texts. Although Transformer-based models such as BERT have improved contextual representation learning, they rely on large-scale labelled data and require extensive task-specific fine-tuning, which limits their adaptability and scalability in dynamic application settings. To overcome these limitations, we propose ELAS-RHP, a novel instruction-tuned framework grounded in Large Language Models (LLMs) that explicitly aligns model behaviour with the characteristics of the RHP task. Specifically, we reformulate review data into prompt–completion pairs and apply Quantized Low-Rank Adapters (QLoRA) to efficiently fine-tune the LLaMA 3 model with reduced computational overhead. By incorporating a few-shot learning strategy, ELAS-RHP enables effective task adaptation under minimal supervision and constrained resources. Empirical evaluations conducted on real-world datasets from Yelp and Amazon demonstrate that our framework consistently outperforms existing baselines across multiple evaluation scenarios. This study provides one of the first empirical investigations into instruction-tuned LLMs for RHP and presents a scalable, efficient and context-aware solution for enhancing review-based information processing in e-commerce environments.

Global population aging has created unprecedented social isolation challenges among elderly populations, with significant negative health consequences. To address this, we propose ELDER-MATCH, a personality-aware expert system that moves beyond conventional interest- or need-based matchmaking. By leveraging AI-powered social robots as an interactive interface, our system helps elderly individuals form not just connections, but suitable and sustainable social relationships based on a deeper understanding of their personality and compatibility. We adopted a two-stage framework: (1) a BERT-based natural language processing stage deriving knowledge from conversational analysis to infer Big Five personality traits, and (2) a K-means clustering stage employing a hybrid knowledge representation to identify compatible social connections based on weighted combinations of personality vectors, interests, and geographical constraints. The social robot provides an intuitive, accessible interface for knowledge acquisition and recommendation delivery, tailored specifically to elderly users with varying technological familiarity. The system was evaluated with 83 older adults across multiple community-based settings. In Stage 2, our unsupervised learning approach identified seven distinct social compatibility clusters, each with specific reasoning rules guiding the recommendation engine. The expert system effectively facilitated meaningful social connections, with 76% of accepted recommendations resulting in ongoing relationships at the three-month follow-up. Beyond interests and needs, personality-aware introductions reduce first-meeting friction and improve trust calibration for older adults. We position ELDER-MATCH as a mediator of human–human ties, and we articulate a “sunset-by-design” principle whereby the system fades as relationships stabilise. Longitudinal assessments revealed significant reductions in loneliness, expansion of participants' social networks, and notable improvements in psychological well-being. These findings demonstrate that a two-stage, personality-aware expert system, coupled with a user-centric interface, successfully bridges technological capabilities and social needs, advancing elderly social wellbeing through responsible AI application within the AI-for-Social-Good paradigm. In practice, personality-aware introductions reduce first-meeting friction and calibrate trust for older adults—benefits that interest-only systems rarely deliver in sustained relationships.

Social media imagery serves as a crucial data source for crisis responders to perceive the evolving crisis situations. The crisis-related information extracted from these images can be used to enhance situational awareness and support decision-making. However, such information provided by data-driven methods is difficult to exploit by model-driven systems, which are widely employed in crisis management practice. This information mismatch caused by the semantic gap undermines the value of social media images in crisis informatics. To address this problem, a metamodel-powered framework for social media image processing is proposed to support crisis response. This framework integrates deep learning techniques, a disaster-specific dataset, information transformation middleware, and a crisis-oriented metamodel. By doing so, it provides ready-for-exploitation information, enabling crisis responders to effectively utilise social media image data. The proposed framework is demonstrated through a case study on the 2018 Aude heave precipitation event and further validated against four additional historical crises. The primary contribution lies in the development of a novel design artefact that follows the design science research paradigm. This study not only addresses the specific information mismatch issue but also offers generalizable design principles applicable to information systems facing similar challenges.