Advances in computational intelligence最新文献

Reliable medical image classification demands two things from deep learning models: high diagnostic accuracy and confidence scores that reflect true uncertainty. Poorly calibrated predictions mislead clinical decision support systems even when accuracy looks acceptable. Hybrid CNN–Transformer architectures capture both local texture and global structural features, but their fusion strategies treat every input identically—a critical limitation, because texture detail and global morphology carry different diagnostic weight across conditions and patients. This work introduces Dual-granularity Adaptive Calibration (DAC), which reframes branch fusion as a per-sample reliability calibration problem. Instead of fixed weights or spatial attention, DAC estimates the reliability of each branch for every input using three complementary signals: predictive entropy, Jensen–Shannon divergence between branch distributions, and exponential moving average feature consistency. A lightweight reliability estimator converts these signals into per-sample fusion weights over branch representations projected to a common feature space. An Adaptive Prototype Classifier (APC) with a dedicated reliability calibration loss then ensures that confidence scores faithfully reflect empirical accuracy. All baselines are reimplemented on identical lesion-disjoint data splits under five independent random seeds to ensure fair comparison; ECE improvements are measured against temperature-scaled post hoc calibration. On HAM10000 (7-class skin lesion, (n = 9015) deduplicated, lesion-disjoint split) and the Kermany Retinal OCT dataset (4-class, (n = 92{,}484)), DAC achieves (93.74 pm 0.29%) and (98.30 pm 0.14%) accuracy whilst reducing Expected Calibration Error (ECE) to (0.031 pm 0.003) and (0.018 pm 0.002)—improvements of 0.031 and 0.030 ECE over the strongest reimplemented baselines (McNemar’s test, (p < 0.001) for both accuracy comparisons, per-seed). A ten-row ablation study confirms each component contributes independently and consistently, and the learned fusion weights show that DAC autonomously prefers local texture on skin data and global structure on retinal data without explicit supervision.

Polycystic Ovary Syndrome (PCOS) is a prevalent endocrine disorder with complex diagnostic challenges due to symptom heterogeneity and limited clinical criteria. This study presents an AI-driven diagnostic framework integrating tabular deep learning, causal inference, and interpretability techniques to enhance PCOS diagnosis. To the best of our knowledge, this is the first study to jointly leverage state-of-the-art tabular models—TabNet, FT-Transformer, and TabPFN—alongside heterogeneous ensemble methods in the context of PCOS diagnosis. Our approach achieves competitive performance, with the Bayesian Neural Network (BNN) attaining 93.83% accuracy on the test set and demonstrating better probability calibration (ECE = 0.02). Key contributions include a Causal-Neural Boosted Ensemble (CNBE) trained on causally significant features via a Causal-Gated Structural Regularization (CGSR) mechanism, and few-shot learning strategies for low-data adaptability. Interpretability is ensured through SHAP analysis and intrinsic attention mechanisms. Causal inference reveals clinically actionable insights, such as the strong causal influence of metabolic and lifestyle factors. The framework contributes to Trustworthy ML principles by providing transparency and uncertainty quantification, offering a foundation for future clinical decision support.

The adoption of photovoltaic solar panels (PVSP) in rural areas remains uneven despite sustained policy support, indicating the need for predictive frameworks that can reliably inform decentralized energy planning in geographically complex regions. This study develops an interpretable and robust machine learning framework for predicting rural household willingness to adopt PVSP. This paper proposes an explainable, gray level co-occurrence matrix (GLCM)-optimized nested stacking ensemble model applied to survey data from rural households in Jammu and Kashmir. Texture-based GLCM descriptors are used to optimize socio-economic variables by capturing latent inter-feature dependencies, while a multi-level nested stacking technique integrates heterogeneous base learners to enhance generalization and prevent information leakage. The concept of model-agnostic explainability was used by employing the aggregation of global local interpretable model-agnostic explanations (LIME) interpretation of adoption factors among multiple households. From the experimental results, it is clear that good generalization has been achieved with an accuracy of 91.38%, F1 score of 91.87%, and ROC-AUC of 97.05%. These results show statistical significance compared with baselines using McNemar’s test (p < 0.05). Factors related to existing installation of PVSP, promotional marketing awareness drives, awareness about government schemes, climatic conditions, and perception were found to be the most important and stable drivers of adoption willingness from the explainability analysis conducted. The proposed framework offers a relevant tool for identifying adoption barriers and designing targeted interventions, contributing to evidence-based planning for decentralized solar energy transitions in rural and geographically constrained regions.

Ensuring that examination questions are pedagogically sound and aligned with the curriculum remains a key challenge in automated assessment. Most existing approaches focus on answer correctness or surface-level textual similarity, overlooking whether questions genuinely reflect syllabus content, require appropriate reasoning, or maintain structural coherence. This study presents SyllabiQA-RSNet, a syllabus-driven framework for automatically evaluating exam question quality. The model combines DeBERTa—large contextual embeddings with graph attention-based syllabus reasoning, capturing dependencies across syllabus sections and multi-step logical connections between questions and course content. Robustness scoring is formulated as a multi-task problem, with the model simultaneously predicting answers, generating reasoning chains, and producing continuous robustness scores. Evaluation on the SyllabusQA benchmark, comprising 5078 question–answer pairs from 63 university syllabi across STEM and non-STEM fields, demonstrates that SyllabiQA-RSNet substantially outperforms baseline models. It achieved a Pearson correlation of 0.79 and Spearman correlation of 0.76 with reference robustness scores, compared to 0.64 for DeBERTa-only and 0.58 for RoBERTa, while reducing RMSE to 0.082. The framework also enhanced answer prediction, particularly for reasoning-intensive questions, improving Exact Match scores by up to 7.2 points and producing higher-quality reasoning chains (+ 8.3 BLEU-4, + 8.8 ROUGE-L). Ablation studies confirmed the necessity of both syllabus graph reasoning and explicit robustness supervision, with their removal causing up to an 11-point drop in robustness correlation. These results indicate that explicitly modeling syllabus structure is critical for reliable, interpretable assessment of question quality, offering insights beyond simple answer correctness and enabling automated evaluation that aligns closely with pedagogical objectives.

Developers and marketers of artificial intelligence (AI) powered innovations have pushed technology to the forefront, well ahead of user experience and ethical implications. This exploratory study used qualitative semi-structured interviews (N = 21) and then quantitative analysis (N = 629) to predict user motivators and detractors to AI adaptation. Fifty-eight percent of respondents in this study did not agree that AI will benefit our society. There was a significant difference those identified as highly feminine and the highly masculine. Those politically conservative were less likely to support AI initiatives than those politically liberal. Independent variables were factored into latent themes “Ethical Outcomes”, Social Media Risk vs. Reward”, and “Governmental Constraint.” Ethical Outcomes was the key significant predictor of “trust AI with my personal and business activities.” The effect of Social Media Risk was marginal but significant. Logistic regression with the dependent variable “overall, artificial intelligence will benefit our society” once again captured Ethical Outcomes. The concern for social media implications was 19% more likely to be considered when pondering beneficence of AI in society. Governmental Constraint was 70% more likely to occur. When AI is considered as a future religious entity, users perceive a decrease in benefit to society. A hierarchical decision tree with the target variable “trust related to AI” illustrates users first consider ethical implications of AI in society, and then face dilemmas of copyright infringement and effect on political elections.

Cardiovascular diseases (CVDs) are the leading cause of global mortality, necessitating efficient and accessible diagnostic tools for early detection. While deep learning models have demonstrated high accuracy in arrhythmia classification, their deployment in resource-limited clinical settings remains challenging due to computational constraints. This study addresses this gap by developing and validating lightweight deep learning models specifically optimized for edge device deployment, enabling real-time ECG analysis in remote and under-resourced healthcare environments. Utilizing the MIT-BIH Arrhythmia Dataset, which includes 48 half-hour ECG recordings from 47 individuals, we trained and evaluated 2 models: a standard Convolutional Neural Network (CNN) and a computationally efficient Depth-Wise Separable Convolutional Neural Network (DWCNN). Preprocessing involved segmenting ECG recordings into individual heartbeats and addressing class imbalance through downsampling, resulting in a balanced dataset of 21,186 images across 6 arrhythmia types. The DWCNN achieved competitive diagnostic performance (precision: 0.97, recall: 0.97) while utilizing 75% fewer parameters than the standard CNN (8.45 M vs. 34.08 M) and 65% less memory (8.4 MB vs. 24.1 MB). Critically, we demonstrate successful deployment on the Intel Neural Compute Stick 2 (NCS2), a resource-constrained edge device, achieving inference times of 4.09 ms/241.91FPS on CPU and 9.05 ms/109.79FPS on the NCS2 platform. This practical demonstration of real-time arrhythmia classification on low-power edge devices represents a significant advancement toward accessible cardiac diagnostics in point-of-care settings, remote monitoring applications, and resource-limited healthcare facilities where centralized computing infrastructure or specialist expertise may be unavailable.

Diabetic retinopathy (DR) stands out as one of the most significant causes of treatable visual impairment and blindness worldwide. Hence, early detection coupled with timely intervention is crucial to prevent the disease from further progression. However, manually detecting DR through the use of retinal images is highly time-consuming, subjective, and often inaccessible in resource-limited settings. This research introduces a groundbreaking automated system for DR detection and severity classification. First, we collected the corresponding data and applied advanced preprocessing techniques such as resizing image, normalization, Gaussian blur, contrast limited adaptive histogram equalization (CLAHE), and image blending to improve model performance. Furthermore, we employed powerful deep learning (DL) architectures such as VGG, EfficientNet, DenseNet, ResNet50, and transformer models such as vision transformer (ViT), and swin transformer for feature extraction. The resulting features were then classified using robust machine learning algorithms, ensemble models, CNN models, and transformer-based models, including decision tree (DT), K-nearest neighbor (KNN), logistic regression (LR), support vector machine (SVM), random forest (RF), AdaBoost, and XGBoost, VGG, EfficientNet, DenseNet, ResNet50, vision transformer, swin transformer, and proposed a new hybrid technique that integrates KNN with random forest (KNN-RF). We tested the system with the appropriate key metrics: accuracy, precision, recall, and F1-score. Our findings show that the hybrid KNN-RF technique outperforms the others. Results point to promising possibilities for our approach in providing precise, scalable, and cost-effective diabetic retinopathy diagnoses in resource-scarce settings. This study emphasizes the importance of artificial intelligence in revolutionizing healthcare diagnostic processes and tackling essential global health issues.

Transfer function plays a crucial role in discretizing metaheuristic algorithms to solve combinatorial optimization problems. However, existing transfer functions not only have few classes, but their design methods also rely too much on the curve shape. Kepler optimization algorithm (KOA) is a novel metaheuristic algorithm that performs well in solving optimization problems on continuous domains, but cannot be directly applied to solve combinatorial optimization problems on discrete domains. In order to design more transfer functions and solve combinatorial optimization problems by KOA, this paper first proposes a practical method to construct transfer functions. From this, two new classes of transfer functions are given: A-shaped transfer functions and B-shaped transfer functions. Then, based on the transfer function, the first discrete Kepler optimization algorithm (DKOA) suitable for binary optimization problems is proposed. To verify the practicality of the new transfer functions and the efficiency of DKOA, DKOA is used to solve 0–1 knapsack problem and knapsack problem with a single continuous variable, respectively. Comparison with existing transfer functions and the state-of-the-art metaheuristic algorithms for solving these two problems shows that DKOA using A-shaped and B-shaped transfer functions is more competitive in terms of the ability to obtain optimal solutions, average performance and stability. This shows that the proposed new transfer functions are very practical, and the DKOA based on them is an effective metaheuristic algorithm for solving binary optimization problems.

Projecting China’s producer price index (PPI) for production materials yields early signals of inflationary pressures and cost dynamics influencing both national economic stability and international supply networks. Reliable PPI forecasts equip policymakers, market participants, and firms with the information needed to refine monetary policy, pricing decisions, and resource allocation. This study proposes an innovative forecasting architecture based on Gaussian process regression (GPR), whose hyperparameters are estimated via a Bayesian inference procedure, enabling the model to adapt in real time to latent market fluctuations and previously unobserved structural shifts. By integrating these evolving characteristics, our approach more accurately captures changes in China’s PPI trajectory. The empirical analysis relies on a monthly dataset spanning October 1996 to February 2025, covering multiple waves of regulatory reform, industrial evolution, and macroeconomic transformation. Validation is performed over an out-of-sample period from June 2019 through February 2025, producing a relative root mean square error of 0.1120%, a root mean square error of 0.1131, a mean absolute error of 0.0832, and a correlation coefficient of 0.99984. To the best of our knowledge, this represents the first application of a Bayesian-inference-parameterized GPR model to forecast China’s PPI for production materials. Beyond advancing the theoretical discourse on machine-learning-based price prediction, the methodology provides a flexible analytical framework applicable to analogous macroeconomic time-series forecasting challenges.

Automated guided vehicles (AGVs) are intelligent robotic systems that play a crucial role in applications, such as transportation, food delivery, and medical emergencies. One of the primary challenges in AGV deployment is achieving precise navigation to ensure task reliability, safety, and efficient battery consumption along predetermined routes. Vision-based methods have gained significant attention for their high performance in robot navigation. However, selecting the most effective algorithm with minimal sensor use remains an active area of research. This study introduces a novel and efficient fusion-based navigation method, termed the Extended Fusion 1 Method (EFM1), which integrates data from camera and infrared (IR) sensors. The system leverages feature-based algorithms, such as SIFT, ORB, FAST, SURF, BRISK, and BRIEF, to improve path-tracking accuracy. The main objective of this fusion approach is to enhance navigational precision and identify the most suitable algorithm for robust AGV path-tracking. The EFM1 method is simulated and validated using the CoppeliaSim (V-REP) simulator, incorporating the real-world dimensions of the previously developed AGV model, Hongma, via Python API. The simulation evaluates six feature-based algorithms across five distinct path types: circular, elliptical, spiral, figure-eight, and custom path. Performance is assessed in terms of maximum achievable speed, minimal path-tracking error, body orientation accuracy, and simulation time. Statistical analysis, including inferential techniques and post hoc tests, is used to interpret the results. The experimental findings demonstrate that the proposed EFM1 algorithm outperforms traditional vision-only approaches in effectively tracking all five path types, confirming its potential for reliable and efficient AGV navigation. The proposed EFM1 algorithm integrates vision and IR sensors, enhancing AGV navigation accuracy by up to 84% while using minimal sensors. It outperforms previous methods by delivering up to 22.2% faster simulation times, significantly reduced error rates (for example, ORB error decreased by 93%), and increased speed across five test paths. The comparison between investigated methods shows that FAST excels on dynamic paths with notable speed improvements, while SURF performs better on complex trajectories as confirmed by statistical analysis. Although EFM1 improves most metrics, body orientation changes increased, indicating a trade-off between agility and movement stability.