Weakly supervised semantic segmentation (WSSS) provides a practical alternative to dense pixel-level annotation, which remains complex and costly in large-scale agricultural monitoring. This work introduces Student-CAM with Chromatic Priors, a fast WSSS framework for agricultural imagery. The central idea is to use Grad-CAM not as a final inference-time explanation, but as a method-level teacher signal generated from the classification branch using only image-level labels. This chromatic-prior-refined teacher signal is distilled into a feed-forward Student-CAM head, allowing class-specific activation maps to be produced at inference without Grad-CAM backpropagation. In this formulation, the chromatic prior constrains activations to plausible vegetation regions, while class-specific Student-CAM responses determine semantic labels. In inference, the Student-CAM module produces class-specific heatmaps in a single forward pass, achieving a 5.5× speedup over gradient-based Grad-CAM. The chromatic prior is applied again to suppress background activations, and region-level voting yields the final segmentation mask. Evaluated on agricultural field imagery, the framework achieves a mean IoU of 76.04
The generalization performance of deep neural networks is often compromised by their tendency to converge to sharp minima and their sensitivity to noisy labels. In this work, we propose SLAW (Sharpness- and Loss-Adaptive Weighting), a unified training framework that synergistically integrates the established concepts of label smoothing, landscape-aware regularization, and sample reweighting to resolve these interconnected problems. The first component, Sharpness-Adaptive Label Smoothing (SALS), adjusts the strength of label smoothing based on a computationally efficient first-order proxy of local sharpness. SALS applies more stronger regularization during training in steep, high-curvature areas of the loss landscape, and gradually reduces the intensity of smoothing as the training stabilizes and the model approaches flatter minima. The second component, Loss-Adaptive Reweighting (LAW), identifies statistically anomalous samples within each mini-batch by analyzing the standardized sample loss values and thereby reduces their influence in the parameter updating process. Unlike several approaches for noise-robust training, LAW does not involve any previous knowledge about noise characteristics, and does not depend on noise estimation rates. Experiments on CIFAR-10, CIFAR-100 and MNIST demonstrate that SLAW has the same performance as conventional training algorithms with clean data and exceptional performance against noisy labels. Under moderate noise (20
The growing share of renewable energy, especially solar photovoltaic (PV) and wind, increases the need for reliable power production forecasts beyond the very short term. Medium-term horizons (here defined as forecasts of at least 24 h ahead) are important for planning tasks such as market bidding, reserve sizing, and operational scheduling, but research in this area is diverse and difficult to compare due to differences in horizons, datasets, and evaluation settings. This paper presents a systematic literature review of machine learning approaches for PV and wind production forecasting with horizons ≥ 24 h. Following PRISMA guidelines, a Scopus search using a broad query (2978 results), followed by a refined query (69 results) and a three-stage screening process (title, abstract, and full text), resulted in 31 included studies. We extract and synthesize evidence on (i) forecast horizons and setups, (ii) model families and hybrid designs, (iii) data sources and the use of measured versus forecast weather inputs, (iv) uncertainty handling (e.g., quantiles, intervals, and scenarios), and (v) links to decision-making contexts. In addition, we propose a taxonomy to classify studies along these dimensions and support comparison across different evaluation settings. The results show that most studies focus on day-ahead and week-ahead evaluation, while 2-week and longer horizons are less common. Hybrid and ensemble approaches are widespread, while transformer-family methods appear in a smaller but visible subset. Most papers rely on measured production data, whereas fewer explicitly use forecast weather (e.g., Numerical Weather Prediction) inputs, even though weather-forecast uncertainty strongly influences multi-day performance in realistic settings. Only a limited number of studies connect forecasts to operational decisions, indicating a gap between accuracy-focused evaluation and decision value. The review highlights the need for more consistent benchmarking, more realistic weather-input evaluation (and, where relevant, other operational exogenous inputs such as curtailment), and more standard uncertainty reporting for medium-term renewable forecasting.
Sixth-generation (6G) communication systems are increasingly being envisioned as learning-enabled infrastructures. However, much of the existing literature treats Big Data and machine learning (ML) as external tools applied to networking problems. This survey adopts a system-level perspective and argues that in 6G systems, data generation and learning are intrinsically coupled with network operations. Observations arise endogenously from communication, control, and architectural decisions, rendering classical assumptions, such as independent sampling, stationarity, and separation between training and deployment, invalid. This work synthesizes Big Data and ML under realistic 6G operating conditions. It characterizes data as a native byproduct of network dynamics rather than a static resource, analyzes the constraints imposed on learning by non-stationarity, partial observability, latency, and resource limitations, and examines the learning behavior within closed communication and control loops. Particular emphasis was placed on edge–cloud data locality and learning placement as determinants of statistical coherence and adaptation timescales. This study further examines the evaluation and benchmarking practices, identifies systematic mismatches between commonly used protocols and the operational 6G reality, and highlights the recurring failure modes induced by these abstractions. It does not aim to catalog learning algorithms or architectural variants but to examine the system-level conditions under which data-driven learning remains valid, stable, and interpretable within operational 6G networks. By reframing Big Data and ML as endogenous system functions of 6G networks, this survey provides a coherent foundation for understanding, evaluating, and designing learning-enabled 6G systems beyond algorithm-centric paradigms.
Stress-related health problems drive an urgent need for sensitive, real-time monitoring devices capable of capturing psychophysiological stress dynamics in natural settings. This study presents a stress detection system that uses deep learning, combining synchronized physiological signals and survey data from university students who experienced controlled stress through psychological sessions and examinations. We analyzed two model families with differing complexities: a deep neural network (DNN)-based model for static physiological patterns and a long short-term memory (LSTM)-based model designed to capture temporal dependencies in biosignals. The DNN (86.42 μ = 2.2, males: μ = 2.7, p < 0.05) and smoking (smokers: μ = 3.7, nonsmokers: μ = 2.2, p < 0.05). Academic elements also impacted stress perception. The stress scores were in accordance with self-reported measurements, which validated the model.
This paper develops a general affine-to-linear transformation and distribution theory for quadratic-loss regression that analytically links affine (with-intercept) and homogeneous (no-intercept) specifications using only summary statistics. Methodologically, we treat a broad class of quadratic-loss problems, including ordinary least squares, ridge regression, and weighted least squares, and show that the no-intercept estimator and its covariance can be expressed in closed form as functions of the with-intercept coefficients, empirical means, and Gram matrices, without revisiting the raw data. We then study three implications of this summary-statistics-only transformation principle: (i) in federated learning, the no-intercept estimator can be recovered from global sufficient statistics by a server-side affine-to-linear transformation, avoiding a second no-intercept run of the federated learning algorithm. We quantify the associated computational and communication costs, compare the one-shot summary-statistics approach with iterative federated optimization, and show numerically that it agrees with the centralized no-intercept estimator up to floating-point precision under non-IID client heterogeneity and in high-dimensional regimes; (ii) in simple linear regression, the transformation clarifies precisely when slopes from with- and without-intercept models have opposite signs (a regression reversal driven by the intercept constraint and leverage) and we characterise its structural analogy with, and distinction from, Simpson’s paradox; and (iii) in biological applications where only aggregate results are available, such as some studies on human proportions and metabolic scaling, the method enables reconstruction of scientifically meaningful through-the-origin fits and their uncertainty directly from published summaries.
Attributed graph clustering or community detection which groups attributed nodes to clusters is a challenging problem and has various applications. The main issues during the clustering of attributed data are obtaining the smoothness of attributes and considering the deep geometrical structure of the graph. Considering these issues simultaneously is one of the main problems that has been neglected. Another problem that could enhance the results is updating the attributes such that the graph topology is close to the data structure. In this paper, we introduce a joint learning framework for updating attributes to be smoother via the deep structure of the graph which is obtained by high-order proximities. Meanwhile, the attributes are updated so that their induced geometrical structure is similar to the graph structure. Experiments on real-world datasets such as Cora, Citeseer, Wiki, and LargeCora demonstrated the effectiveness of the proposed method for community detection. Notably, our method achieved approximately 25% higher accuracy on LargeCora compared to the second-best method, and an improvement of 4.5% on Citeseer. Furthermore, it outperformed all compared methods in F1-measure for Cora, Citeseer, Wiki, and LargeCora, securing the second-best position for Pubmed.
Single-cell RNA sequencing (scRNA-seq) enables high-resolution profiling of cellular heterogeneity, yet its high dimensionality, sparsity, and prevalent dropout events pose significant challenges for unsupervised cell-type clustering. To address these challenges, we propose SaDGAE, an unsupervised deep graph autoencoder (GAE) framework that jointly models gene expression patterns and cell–cell relationships. SaDGAE first employs a ZINB-based autoencoder to generate dimensionally reduced representations that mitigate technical noise while preserving biological variation. Based on these representations, we introduce a self-adaptive cell graph construction mechanism, which dynamically refines cell–cell relationships during training through joint optimization with embedding learning and clustering, to capture subtle and context-specific connectivity. Built upon this adaptive graph, a doubly enhanced GAE (DeGAE) is designed to encode both structural and feature-level information, reconstructing the cell graph from latent embeddings and decoder-recovered features to provide a more informative graph reconstruction signal and more discriminative embeddings. Finally, clusters are inferred through a self-optimizing clustering module that aligns learned embeddings with target distributions in a self-supervised manner. Extensive experiments on benchmark scRNA-seq datasets, together with case studies and an additional large-scale scalability study, demonstrate that SaDGAE achieves strong and competitive clustering performance, yielding biologically interpretable clusters and accurately recovering known marker gene patterns.
Deep reinforcement learning (DRL) has emerged as a powerful framework in medical image analysis, enabling sequential decision-making for tasks such as annotation, diagnosis, localization, and segmentation. Despite numerous studies in this domain, there is still a lack of systematic understanding of methodological choices, data types, and task-specific approaches. In this survey, we provide a structured analysis and multi-dimensional taxonomy of DRL applications in medical imaging, highlighting the advantages and limitations of existing approaches. In particular, we organize the literature based on multiple dimensions, including medical data type, task category, and DRL techniques, offering an integrated perspective that supports better understanding and comparison of current methods. This work aims to provide practical insights and guidance for future research and development in DRL-based healthcare systems.
Stance detection aims to identify an author’s stance toward a target topic. Although recent prompting-based methods enable zero- and few-shot stance detection with large language models (LLMs), they often require extensive prompting or costly use of multiple LLM agents, as well as large labeled training datasets. We present MINE (Meta-prompting and In-context N-shot Examples), a simple, light-weight, training-free framework that first creates a target-specific LLM stance agent through meta-prompting, and then augments it with a small number of semantically retrieved examples. While meta-prompting and retrieval-augmented in-context learning have been studied independently, these components have not been combined specifically for stance detection in a training-free pipeline to the best of our knowledge. We also provide a systematic empirical study that quantifies the effects of the example retrieval strategy, sentence encoder choice, demonstrator pool size, and the number of examples on the performance of this hybrid methodology. We show that MINE achieves competitive performance when deployed as either a zero-shot or few-shot stance detector, matching or exceeding existing prompting-only approaches on SemEval-2016 and P-Stance while using fewer prompts and demonstrations. MINE remains effective with minimal labeled data (e.g., 10
Multi-view data have become prevalent across complex application scenarios, where heterogeneous views provide complementary descriptions of the same objects. Effectively exploiting such complementary information while suppressing redundancy remains a central challenge in multi-view clustering. In this paper, we propose a Concept Factorization-based Multi-View Clustering framework (CFMvC) that unifies diversity preservation, orthogonality enforcement, and manifold learning for multi-view clustering. Specifically, CFMvC simultaneously integrates four key components: (i) cross-view diversity regularization to capture complementary information across views, (ii) within-view orthogonality constraints to ensure discriminative cluster structures, (iii) graph Laplacian manifold regularization to preserve the intrinsic geometric structure of the data, and (iv) an adaptive view-weighting scheme to automatically determine view importance. These components collectively construct a robust consensus representation that is subsequently partitioned via k-means clustering. Extensive experiments on ten benchmark datasets demonstrate that CFMvC consistently outperforms state-of-the-art algorithms, achieving average improvements of 6.25
Large language models (LLMs) face significant limitations in specialized algorithmic domains because they do not reliably follow formal specifications and often produce non-executable or semantically incorrect solutions. This limitation is particularly critical in the field of pattern mining, where consistent adherence to algorithm definitions, parameter semantics, dataset formats, and library-specific interfaces is essential for producing valid results. To address this challenge, we present PAttern MIning-Generative Pre-trained Transformer (PAMI-GPT), a computational framework that integrates conversational AI with pattern mining expertise through a curated Domain Knowledge Base derived from the PAMI library and its verified algorithmic specifications. The proposed framework combines natural language understanding with domain-specific algorithm guidance and executable workflow templates at inference time. Compared to standard LLM prompting across evaluated models, PAMI-GPT achieves markedly higher semantic correctness (+ 57.0 pp), executability (+ 75.5 pp), and reduction in hallucinated APIs (− 50.4 pp). We further perform component-level ablation and inter-annotator agreement analysis to evaluate the contribution of structured retrieval and instruction-guided generation. Experimental evaluation on 300 realistic pattern mining tasks demonstrates consistent improvements in algorithm configuration accuracy, parameter compliance, and dataset specification correctness across zero-shot, one-shot, and few-shot settings. These results highlight the effectiveness of inference-time domain grounding for enabling more reliable and reproducible pattern mining workflows without requiring model retraining.
Understanding dynamic customer behavior is essential for effective segmentation in evolving transactional and relational environments. Traditional clustering methods and static knowledge graphs capture only limited, fixed patterns, while deep clustering approaches often require full retraining and neglect relational and contextual information. To address these challenges, we propose IncDEC, a framework that integrates DKG with BERT-based semantic embeddings. IncDEC models evolving relationships between customers, products, and transactions, enabling rich structural and contextual representations. By incrementally updating cluster assignments and centroids in the latent space, IncDEC preserves cluster quality while adapting to new customer behaviors without retraining. Experimental results demonstrate its effectiveness, scalability, and adaptability for real-world dynamic customer segmentation.
Non-causal feature selection methods may introduce errors in assessing feature relevance because they rely on a single selection criterion, which can lead to spurious correlations. The current causality methods of feature selection consider only the features that compose the Markov blanket of one target variable (class) as relevant. That may lead to the disregard of other relevant features. We present a new algorithm, LUISA, based on causal relationships. This algorithm expands the Markov blanket to include indirect ancestors for a specific target variable or class. We test the algorithm on five perspectives using synthetic and real-world datasets. The experiments show that LUISA is representative and that it performs satisfactorily under uncertainty. In addition, the learning models demonstrated promising results when using the reduced dataset compared to the complete dataset and various feature non-causal and causal selection methods.
Learning Chinese grammar is a difficult endeavor to pursue because of the combination of linguistic properties. The traditional rule-based approach in pedagogy tends to be rigid, and because it cannot facilitate individualized needs or scaffold the learning experience, it can be inefficient and demotivating. In this paper, we present a contribution to addressing these shortcomings in pedagogy through a hybrid natural language processing (NLP) and deep reinforcement learning (DRL) model to derive an adaptive teaching algorithm for Chinese grammar. This model incorporates NLP for learning grammar and DRL for dynamic choice of teaching strategy based on learner profiles, with personalized feedback and real-time adjustment of learning difficulty. Distinct from traditional AI algorithms that are fixed, the new model incorporates proximal policy optimization (PPO) and experience replay, which contributes to the levels of adaptability and stability of learning. Experimental results demonstrate that the proposed DRL-based adaptive teaching system reduced the average time-to-mastery from 12 to 4 sessions (66.7
Bayesian Networks (BNs) provide an interpretable probabilistic framework for modeling uncertainty in complex environmental systems. This study proposes a data-driven Bayesian Network (BN) to represent probabilistic relationships among meteorological variables and precipitation in Rio de Janeiro, Brazil, using hourly data from telemetric stations (2002–2024). The methodology combines physically motivated feature engineering, PCA-based assessment of redundant temporal-variation features, and structure learning under meteorology-informed constraints to ensure temporal and physical coherence. Structural constraints based on meteorological knowledge ensured temporal and physical coherence. Among the evaluated models, the Hill-Climbing algorithm with the K2 criterion achieved the best predictive performance in terms of accuracy, specificity and calibration. However, for interpretative analysis, the Hill-Climbing algorithm with the Bayesian Information Criterion (BIC) criterion was selected due to its more parsimonious structure, which aligns better with physical reasoning. The learned structure indicated that rainfall depends mainly on short-term atmospheric conditions, particularly high humidity, low solar radiation, and moderate wind speeds at one-hour lags. Validation confirmed consistent cross-validation results and physically coherent patterns associated with convection and moisture transport. By explicitly representing conditional dependencies, the proposed BN improves interpretability over black-box methods and supports probabilistic reasoning for short-term precipitation assessment. The findings highlight the potential of BNs as transparent and physically consistent tools for weather modeling and environmental decision support. Furthermore, BNs proved to be a sustainable modeling approach, characterized by low computational demand and minimal energy consumption during the training phase.
Traditional sequential recommendation models often rely solely on user interaction histories, neglecting rich contextual signals like visual content and ratings. This narrow focus limits their ability to capture the full complexity of user preferences and fails to account for behavioral uncertainty. To address these challenges, we propose AVMSR, an Attentional Variational Hybrid CNN-LSTM model for sequential recommendation. AVMSR integrates multi-modal data by extracting visual features using CNNs and encoding user ratings through dense embeddings. These inputs are fused at each time step and processed by an LSTM to model temporal dynamics. An attention mechanism highlights influential historical interactions, while a variational inference component captures latent uncertainties via the reparameterization trick and KL divergence. This design enables a more expressive and robust user preference modeling. Experimental results on real-world datasets (MovieLens, Amazon, and Yelp) demonstrate that AVMSR significantly outperforms state-of-the-art approaches on Recall, NDCG, and MRR, validating the impact of multi-modal fusion and probabilistic reasoning.
The global threat of infectious diseases is being exacerbated by the issue of antimicrobial resistance (AMR). Drug resistance is growing, thus investigating antibiotic resistance strains in E. coli and urine tract infections (UTIs) is essential for effective diagnosis. Although next-generation sequencing (NGS) has made whole genome sequencing (WGS) widely available and used in AMR studies, it is still challenging to identify bacterial strain that are resistant using sequencing data. Rapid strain detection by deep learning (DL) in combating infections has reduced fatality rates. Introducing ARGai 3.0, a fusion of artificial intelligence (AI) models for resistant E. coli strain detection that is both generalized and streamlined. The ARGAi 3.0 framework for antibiotic-resistant strain classification represents a developing integration of quality control with AI models. Several classification metrics, including sensitivity, specificity, accuracy, precision, ROC-AUC, and F1-score, were assessed using ARGai 3.0 through the use of cross-validation. We hypothesized that, in comparison with baseline AI models, ARGai 3.0 would have superior capability in identifying strains of antibiotic resistance. In comparison with traditional AI models, our ARGai 3.0 demonstrates a 10
The rapid digital transformation of the tourism industry has led to the widespread adoption of artificial intelligence (AI), Internet of Things (IoT), and data-driven intelligent services. Despite these advances, safety management in smart tourism environments remains largely reactive and lacks real-time predictive analytics. This paper presents a hybrid edge-cloud AI-based data analytics framework for real-time predictive safety alert generation in smart tourism ecosystems. The proposed framework combines low-latency edge-level analytics with cloud-based large-scale learning to enable timely safety decision-making and continuous model adaptation. Spatiotemporal safety risks are modeled using a Graph Convolutional Network–Long Short-Term Memory (GCN–LSTM) architecture, while anomalous safety events are detected using an Autoencoder–GRU model. The framework supports the fusion and analysis of heterogeneous data sources, including IoT sensor streams, GPS trajectories, environmental conditions, and contextual signals, enabling comprehensive real-time safety analytics. Predictive inference is performed at the edge to ensure rapid response, while cloud-based analytics facilitate model retraining and performance optimization over time. The proposed approach is evaluated using the CrowdFlow and METAR datasets. Experimental results demonstrate that the hybrid analytics framework outperforms conventional cloud-only and edge-only approaches in terms of prediction accuracy, response latency, and energy efficiency. The findings highlight the effectiveness of hybrid data analytics for proactive safety management and demonstrate the potential of the proposed framework to enhance safety intelligence and operational efficiency in smart tourism ecosystems.
Large language models (LLMs) increasingly provide health information, yet their performance in multilingual, middle-income settings remains under-evaluated. We developed a framework to benchmark LLM-generated health information across languages while considering factors beyond accuracy. In a convergent mixed-methods study, ChatGPT, DeepSeek, and Gemini answered ten vaccine-related questions in English and Malay. Outputs were compared with guideline-derived reference responses and evaluated by a multidisciplinary expert panel for accuracy, completeness, misinformation, readability, and overall preference, with qualitative commentary. Across both languages, experts rated LLM responses as highly accurate and complete, with low misinformation, and generally preferred them to the official responses. ChatGPT received the highest proportion of top rankings (50
