Automated essay scoring (AES) has attracted increasing attention as a scalable solution for high-stakes language assessments such as IELTS. However, many existing AES systems—particularly those based on large language models (LLMs)—exhibit limited construct alignment and unstable agreement with human raters, especially in domain-specific assessment contexts. This study proposes a feature-calibrated framework for LLM-based essay scoring. Sixteen linguistically and discourse-informed subfeatures, explicitly aligned with the official IELTS writing rubric, are encoded using pointwise mutual information (PMI) and integrated as a calibration layer on top of a lightweight Llama-2-7B architecture. This design aims to strengthen construct representation and improve rater alignment while preserving the semantic modeling capacity of LLMs. Experimental results show that the calibrated model achieves high agreement with expert ratings, reaching a quadratic weighted kappa (QWK) of 0.9321 and outperforming previously reported LLM-based AES approaches, including a GPT-4-based study. Additional experiments demonstrate robust performance under low-resource conditions, with competitive agreement obtained using as few as 500 training samples. Overall, the findings suggest that theoretically grounded calibration provides an effective pathway for improving the construct alignment, robustness, and scalability of LLM-based AES systems in high-stakes writing assessment.
Ensemble learning improves the accuracy of the neuro‐fuzzy systems but compromises their interpretability because of the separate, independent rule bases. In this paper, we propose SimBAC‐NFS, which combines the variance reduction of bagging with the residual learning of boosting to produce a single interpretable model and rule base. The model assesses similarity and identifies rules describing similar input regions across all learners, merges them into consistent membership functions and refits the consequent linear parameters on the full training data. SimBAC‐NFS achieves the lowest error on eight of the nine datasets among single and classical fuzzy ensemble models on nine engineering regression benchmarks. Against unconstrained black‐box baselines such as XGBoost, LightGBM and Random Forest, it wins outright on the two datasets with the smoothest, structurally simplest responses. On the remaining seven, it accepts a bounded accuracy gap in exchange for a single rule base that a domain expert can read and audit directly.
The Fourth Industrial Revolution has driven the formation of smart manufacturing—where production processes are highly automated based on real‐time data. A core component of this model is the Industrial Internet of Things (IIoT), which enables devices, such as robots, CNC machines, sensor systems, AI cameras, conveyor belts, and automated guided vehicles (AGVs) to connect into a unified network. These devices not only collect and transmit data but also coordinate processing, decision‐making, and automatically adjust operations to optimize production performance. In the smart factory model, IIoT systems are deployed to collect and transmit status data from production equipment, such as CNC machines, robots, sensors, conveyor belts, and AGVs. This infrastructure is typically built according to a layered architecture, in which IoT devices and the edge layer are responsible for recording signals and status in real‐time, while the fog and cloud layers support data storage and synchronization on a broader scale. Through this continuous connection mechanism, the system can accurately observe the operational status of machinery and workflow at the shop floor, creating a reliable data foundation for monitoring, analysis, and scheduling tasks in smart factories. Integrating IoT into manufacturing helps factories achieve many important benefits, such as (1) reducing cycle time through automated monitoring and coordination, (2) increasing accuracy and product quality through sensors and AI inspection, (3) reducing energy costs through data‐driven operation optimization, (4) increasing self‐recovery capability when the system can automatically detect and adjust to failures, and (5) optimizing workflow with tight coordination between devices. Workflow scheduling plays a role in intelligent coordination among robots, CNC machines, sensors, AI inspection systems, and automated transportation infrastructure, ensuring that each task is executed on the most suitable type of resource, within the system’s capacity limits, at the most appropriate time while contributing to reducing energy costs. This research aims to develop an intelligent scheduling algorithm based on deep reinforcement learning combined with graph attention networks (GATs) to optimize the completion time of workflows in the smart factory model.
Ultrasound imaging is widely used for diagnosis and monitoring because it is safe, portable, cost‐effective and suitable for real‐time clinical assessment. However, speckle noise, acoustic shadowing, attenuation, low tissue contrast and operator‐dependent variability remain challenging for automated prediction and spatial localisation. Although deep learning has improved ultrasound analysis, many models remain difficult to audit clinically. Common post hoc explanation methods, including Grad‐CAM, integrated gradients (IGs), LIME and SHAP, can produce diffuse, unstable or anatomically inconsistent attribution maps whose functional relationship with the underlying prediction is not guaranteed. We present FRAME, an explanation‐by‐design framework in which prediction and explanation are coupled through the same spatial evidence field. A lightweight U‐Net–style encoder–decoder generates dense evidence, and the scalar output is obtained directly by global average pooling of this field. Expert masks provide spatial regularisation during training. FRAME was evaluated on complete held‐out test sets across four ultrasound settings: BUSI breast ultrasound, microultrasound prostate, ulcerative colitis transabdominal ultrasound and FOCUS foetal four‐chamber ultrasound. BUSI was evaluated as diagnostic image‐level classification with spatial explanation, whereas the remaining datasets were treated as anatomy‐centred evidence‐localisation tasks. FRAME achieved Dice scores of 0.901 ± 0.061, 0.893 ± 0.048, 0.831 ± 0.031 and 0.887 ± 0.058 on BUSI, prostate, ulcerative colitis and FOCUS, respectively, with corresponding XAlign scores of 0.916 ± 0.052, 0.910 ± 0.043, 0.852 ± 0.028 and 0.905 ± 0.049. These values exceeded Grad‐CAM, IG, LIME and SHAP, as well as supervision‐controlled and mask‐free baselines. Empirical faithfulness analyses using deletion, insertion, sufficiency, comprehensiveness, evidence masking, model randomisation and perturbation robustness further showed that FRAME evidence was more strongly related to the model output. Component‐wise ablations confirmed contributions from evidence‐field supervision, global average pooling, focal loss, test‐time augmentation and the lightweight architecture. Independent radiologist and clinician assessment also showed high acceptance of FRAME explanations. Analysis of incorrect predictions demonstrated reduced spatial alignment, highlighting that FRAME remains susceptible to incomplete or off‐target evidence in failure cases. Overall, FRAME provides a computationally efficient and empirically validated approach for coupling prediction with anatomically grounded evidence in heterogeneous ultrasound applications.
Background Timely and accurate predictions are essential for responding to infectious disease outbreaks and resurgences. Monkeypox (mpox) has re‐emerged as a global health concern, necessitating effective forecasting tools to guide intervention strategies. This study develops and evaluates a hybrid machine learning (ML) framework to forecast mpox transmission trends in North America. Methods Daily confirmed mpox cases were obtained from Our World in Data. A hybrid forecasting framework was developed by integrating classical time‐series models (ARIMA, ETS, TBATS, and Prophet) with ML algorithms (artificial neural network [ANN], support vector regression [SVR], and random forest [RF]). Multiple hybrid combinations were systematically evaluated to capture both linear and nonlinear transmission dynamics. Model performance was assessed using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and first lag autocorrelation of residuals (ACF1), with residual diagnostics ensuring temporal independence. Results The Hybrid ANN_RF model achieved the highest predictive accuracy (RMSE = 23.73, MAE = 21.53, MAPE = 0.83, and ACF1 = 0.72), outperforming both individual and other hybrid models. Out‐of‐sample forecasts predicted sustained low‐level transmission (1–5 cases/day) through mid‐2026, with no warning or alert signals detected based on defined surveillance thresholds. Conclusion Hybrid ML models demonstrate superior predictive capability compared to individual models for epidemiological forecasting. The proposed ANN‐RF model provides reliable and operationally useful predictions that can inform proactive public health measures, including vaccination planning, early warning surveillance, and resource optimization.
