Expert Systems最新文献

Evaluating open-ended questions is a common and time-consuming task in education. With the continuous advances in natural language processing (NLP), large language models (LLMs) trained on massive datasets can assist in this process. This study evaluates the use of LLMs, complemented by retrieval-augmented generation (RAG), for the numerical grading of open-ended answers of approximately 250 words. We focus on two Spanish-language technical courses and assess general-purpose LLMs. Our results show that RAG improves grading accuracy, achieving reductions in mean absolute error (MAE) of up to 19.47% compared to using LLMs alone, with the best configuration reaching a MAE of 1.19. We also note that LLMs tend to assign high grades, reflecting the dataset's imbalance toward higher scores. This work demonstrates the potential of combining RAG with general-purpose LLMs to evaluate specialised Spanish language content, avoiding the cost and bias of model fine-tuning.

Historical photographs and artworks often suffer from degradation, missing regions, or stylistic corruption due to aging, scanning artefacts, or incomplete archival processes. While recent image inpainting and restoration models achieve plausible visual reconstructions, they often disregard the cultural and temporal context of the content—producing restorations that are visually coherent yet stylistically anachronistic. In this paper, we present TimeBrush, a temporally guided diffusion-based framework for historical image restoration. By conditioning the generation process on explicit temporal prompts (e.g., art period, century), and reinforcing stylistic alignment through a learned style consistency discriminator, TimeBrush faithfully reconstructs missing content while preserving culturally significant visual traits. Our framework integrates a Temporal Prompt Encoder and a Style Consistency Discriminator, allowing restorations to be faithful both temporally and stylistically. TimeBrush improves Style Accuracy by over 7% compared with state-of-the-art baselines, while also having better perceptual quality. These results indicate TimeBrush's promising opportunities for AI-assisted cultural heritage preservation and museum digitization.

High-performance surface defect detection is essential for industrial quality inspection, requiring accurate detection and characterisation of surface defects. While deep learning-based methods have advanced this field, challenges persist due to limited sample availability and variation in defect types. To address these challenges, we propose a new detection framework, namely, the Dual-branch Cross-Diversion Transformer with Spatial Soft Alignment, specifically designed for surface defect detection under data-scarce conditions. First, the dual-branch cross-transformer is leveraged to address data scarcity and enhance defect detection sensitivity through a few-shot pipeline. Furthermore, the Adaptive Activation Downsampling module is proposed to capture coarse-grained structural features while preserving fine-grained defect details, ensuring comprehensive surface defect characterisation. Additionally, a Cross-Diversion Self-Attention mechanism further improves multi-scale feature extraction, critical for accurate detection of diverse defect types. Finally, a Spatial Soft Alignment strategy corrects spatial misalignment between detection proposals and defect categories, reducing detection uncertainty. Through extensive experiments on two benchmark industrial datasets, our proposed architecture achieves superior performance compared to state-of-the-art methods, demonstrating its robustness and accuracy. These results demonstrate the effectiveness of the proposed method and its potential to advance surface defect detection techniques.

Aspect-level sentiment analysis (ALSA) is the process of determining the emotional polarity that people have towards aspects of topics or entities expressed in their opinions. ALSA is increasingly integrated into many practical applications to make them more user-friendly and suitable for users' psychological and emotional trends. Therefore, the performance of ALSA methods is increasingly being studied by scientists for improvement. Various approaches have been proposed for ALSA, the latest of which is Graph Convolutional Networks (GCNs). Although they have performed well, previous GCN-based methods still fail to capture all important features from opinions. This raises the question of whether combining ALSA-based GCNs can improve the ability of previous methods to capture important features. This motivates us to propose the ALSA method based on the Ensemble Graph Convolutional Networks (EGCNs). The objective of the proposed method is to capture features in a manner that is both independent and joint, in order to leverage the advantages of jointly learning features while also benefiting from the strengths of learning features independently. The proposed method includes the following main steps: (i) data representation based on the BERT model; (ii) extracting syntactic, semantic and contextual features based on the ASGCN, ATGCN and ASCNN models, respectively; (iii) combining the extracted feature vectors into a general feature vector based on the fusion mechanism; (iv) sentiment analysis based on the Softmax function. To demonstrate the performance of the EGCNs model, it is experimented on three benchmark datasets and compared with the previous methods before being combined.

Federated recommendation systems aim to provide personalised services in decentralised environments while preserving user privacy. However, under strict privacy constraints and limited local information, existing federated models struggle to capture user-item interactions effectively. This paper proposes FedLGCN, a novel federated recommendation framework that combines Lightweight Graph Convolutional Networks with a privacy-enhanced local training protocol. A hash-based privacy-aware graph augmentation strategy is introduced to enrich each client's local subgraph without disclosing sensitive neighbour information. Additionally, Local Differential Privacy is employed to perturb gradients before aggregation, providing robust protection against inference attacks. The Lightweight Graph Convolutional Networks-based embedding module enables efficient and scalable representation learning with reduced communication and computational costs. Extensive experiments on three public benchmark datasets demonstrate that FedLGCN achieves a well-balanced trade-off among communication cost, recommendation accuracy and privacy protection, consistently outperforming several baseline methods in federated recommendation.

Federated learning has emerged as a powerful paradigm for collaborative machine learning across multiple parties, holding considerable potential for modern industries. However, its inherently decentralised and collaborative nature raises critical concerns about data security and user privacy. Sensitive information—such as user preferences, behaviours, and identities—remains susceptible to inference attacks, revealing the limitations of conventional privacy-preserving techniques in existing federated learning frameworks. To address these challenges, this paper presents GSSFL, a novel federated learning architecture that integrates smart contracts and group signatures to enhance both privacy protection and system trustworthiness. GSSFL enables secure and verifiable data exchange without compromising user anonymity, while its decentralised design encourages broader participation in federated learning processes. Experimental results demonstrate that GSSFL effectively satisfies the demands of privacy-preserving data sharing with minimal performance overhead.

Causal discovery focuses on identifying a target's direct causes and effects (e.g., class label) feature of interest in a Bayesian network (BN). Existing causal discovery primarily includes global and local learning algorithms, which must access the whole feature space before the learning process starts. However, many real-world applications continuously generate features in real-time and demand stream processing of features for just-in-time decision-making. In addition, the local and global learning algorithms either emphasise accuracy or computing efficiency over both. Therefore, to address these problems and handle dynamic high-dimensional feature space, we proposed a novel local causal discovery algorithm based on streaming features that consider improving and balancing both computational efficiency and prediction accuracy denoted as Local Causal Discovery Towards High-Dimensional Streaming Features $��\left({\mathrm{LCD}}_{\mathrm{SF}}\right)�$. More specifically, to attain this objective, $��{\mathrm{LCD}}_{\mathrm{SF}}�$ dynamically integrates $��V�$- and $��N�$-structure to learn the Markov blanket (MB) and simultaneously distinguishes the direct causes (parents) from direct effects (children) and parents–children (PC) from spouses of the target feature. Thus accomplishes the balance between efficiency and accuracy of prediction. The proposed algorithm has been extensively evaluated on 10 benchmark BNs and 10 real-world datasets. The results show that the proposed algorithm outperformed the existing state-of-the-art baseline algorithms. The source code is available at: https://github.com/vickykhan89/LCDSF.

Cardiovascular diseases remain the leading cause of death worldwide, highlighting the need for expert systems that enable continuous and interpretable cardiac monitoring. We present ArmFormer, a knowledge-driven expert system that leverages Transformer-based reasoning for robust cardiac event detection from wearable armband electrocardiogram signals. The model integrates domain-guided multi-scale patch encoding to capture waveform morphology and rhythm dependencies, while local gated Transformer blocks enhance temporal continuity and suppress noise-induced variability. A lead-wise attention mechanism coupled with gradient-based visualisation provides interpretability by highlighting clinically relevant regions such as QRS complexes, P waves, and ST segments. On an in-house cohort of 99 subjects comprising 6211 normal and 10,030 abnormal 10-s segments, ArmFormer achieved 91.66% accuracy, 91.57% F1-score, 91.41% sensitivity, and 97.36% AUC under a subject-exclusive protocol that prevents patient-level information leakage. Compared with convolutional and residual baselines, AUC improved by up to 5%, while floating-point operations and parameters were reduced by 29-fold and 47-fold, respectively, achieving 2.58 ms inference latency per segment. External validation on CPSC2018 and Chapman showed accuracies of 83.65% and 94.69%, with AUCs of 96.69% and 99.38%, respectively. By combining domain-guided encoding, noise-robust temporal reasoning, and interpretable attention, ArmFormer provides a practical and reliable framework for expert-level cardiac event detection in wearable monitoring scenarios.

Although significant progress has been made in multi-view classification over the past few decades, handling multi-view data with arbitrary view missing is still a challenge. To address the challenge of incomplete multi-view classification, we propose a novel framework named Supervised Variational Incomplete Multi-View Classification (SVIMC) network, which completes incomplete multi-view data and performs classification predictions. Specifically, we design a supervised multi-view variational autoencoder for missing view completion, which involves a Product of Experts (PoE) network to obtain the latent joint representation of available views. This representation is then fed into the view-specific decoders of the missing views to generate the imputations. Besides, we jointly optimise the classification network and the missing view completion module, allowing them to mutually promote each other. Moreover, by adopting the GradNorm method, we significantly reduce the difficulty of model training. Extensive experiments have been conducted to demonstrate the effectiveness of our method in terms of classification accuracy, missing view imputation visualisation and ablation study.

Recent advances in deep learning have significantly propelled progress in person re-identification. However, many current solutions often prioritize architectural optimization. Although this approach has led to considerable performance improvements, it may inadvertently overlook the critical challenge of suppressing interference from complex background noise. To explore a path toward addressing this aspect, we propose a Multi-scale Feature Enhancement Network (MSFENet). Our approach includes a spatial-frequency fusion module designed to guide the network's attention toward pedestrian-specific regions. Furthermore, the incorporation of frequency-domain cues is intended to facilitate the capture of fine-grained details, thereby potentially enhancing robustness. We also design a Multi-Granularity Fusion (MGFusion) module to help alleviate overfitting and information loss during feature interaction. Experimental results indicate that MSFENet achieves competitive performance across evaluated tasks on the Market1501 and MSMT17 datasets, as well as under cross-domain settings (MSMT17 → Market1501).