Combining large language models (LLMs) with structured domain knowledge representation (knowledge graph) and software agents appears to have potential for supporting domain-specific, open-ended situation awareness and scenario building. Inherent to these processes is that there is a complex and dynamic phenomenon under investigation, required input data is gathered from both structured and unstructured sources with varying confidence and quality levels, and process execution requires both human-human and human-AI collaboration. In this paper, we present a design, implementation, and partial validation for a data processing pipeline consisting of an LLM-based AI implementation and a collection of software agents, applied to a process of military intelligence. Our contributions are: 1) survey and eight-phase guidance of knowledge representation for LLMs, 2) survey and guidelines of trust in AI application adoption, and 3) implementation of agentic AI data processing pipeline for synthetic observation generation and processing. Our key findings indicate that through structured domain knowledge support, the foundation model is capable of generating outputs subject to domain expert evaluation, and that the utilization of software agents allows for experimenting with potential course-of-action (CoA) scenarios from a present situation. Future work requires more rigorous testing with domain experts to identify improvements needed in the data processing pipeline.
Accurate stratigraphic identification is essential for petroleum-development planning, but class imbalance and similar well-log responses between adjacent strata remain major challenges. This study uses well-log data from nine anonymized wells in the Ordos Basin, covering approximately 1,200–2,400 m and nine stratigraphic units from Yan 9 to Chang 10_1. Seven wells were used for training and hyperparameter optimization, while Wells H and I were reserved for blind evaluation. A two-stage framework was developed by decomposing the nine-class task into three-class and seven-class subtasks using an operational depth constraint within Yan 10_2. Savitzky–Golay filtering was applied, and XGBoost, CatBoost, LightGBM, Stacking, and Att-BiLSTM were evaluated. Three composite features were constructed before SHAP analysis based on petrophysical relationships and exploratory analysis of the training wells. SHAP was used for post-hoc interpretation rather than feature construction or selection. CatBoost achieved the strongest overall performance among the evaluated models, although no single model performed best for every stratum. With an adjacent-stratum boundary tolerance of approximately ±1 m, the framework achieved accuracies of 93.4% and 95.1% for Wells H and I, respectively, with macro-F1 scores of 0.867 and 0.835. These results demonstrate its potential for stratigraphic identification within the study area.
Federated learning in healthcare faces significant challenges under non-independent and identically distributed (non-IID) medical data, including communication inefficiency, unstable convergence, and privacy risks. To address these challenges, we propose MedSparseFL, a sparse and privacy-preserving federated learning framework for healthcare applications. On the client side, MedSparseFL introduces a gradient sparsification mechanism using a lightweight Gradient Score Network (GSN) to estimate the importance of gradient updates. High-contribution dimensions are preferentially selected through probabilistic sampling combined with residual feedback, thereby alleviating the adverse effects of non-IID data on global convergence. Before transmission, the sparse gradients are compressed using Count Sketch mapping and protected via sketch-domain obfuscation and pairwise masking to enhance privacy. Additive homomorphic encryption (AHE) is further applied to the compressed sketches to reduce transmission leakage risks while maintaining computational efficiency. On the server side, global gradient estimates are obtained from aggregated sketches, and the inferred sparse support is exploited to suppress anomalous updates. Experiments on two public medical datasets show that MedSparseFL improves classification performance by 2%, reduces communication overhead, and lowers empirical privacy leakage risks under the evaluated attack settings, with privacy protection positioned at the update level under the stated threat model.
Advances in artificial intelligence have made Multiple Appropriate Facial Reaction Generation (MAFRG) increasingly important in human-computer interaction. However, existing models face practical deployment challenges due to complex structures and high computational costs. This paper proposes a lightweight framework for generating appropriate multimodal facial reactions, comprising two complementary methods targeting structural sparsity and numerical precision, respectively. DyPR-Net dynamically adjusts its sparse structure via adaptive thresholding and a weight regeneration mechanism to restore critical connections, balancing compression efficiency with facial reaction fidelity. GOMD-Net addresses semantic drift and degraded expression quality under low-bit quantization by adaptively allocating bit widths across layers based on inter-layer orthogonality. It further employs a Distribution-Skew-Correction (DSC) strategy combined with a global optimization objective to faithfully reproduce subtle and diverse emotional expressions under extremely low bit-width constraints. Extensive experiments across multiple mainstream architectures on the NoXi and RECOLA datasets demonstrate the effectiveness and potential generalizability of the proposed framework, reducing model size, FLOPs in the tested environment, thereby offering a promising basis toward the lightweight deployment of MAFRG models.
Short-horizon demand forecasting in industrial settings is often characterized by sparse observations, frequent zero-demand periods, and limited historical data at the customer – product level. These conditions pose significant challenges for conventional forecasting models, which treat demand as a continuous process and do not explicitly account for zero inflation. This study proposes a zero-inflated neural forecasting framework that decomposes demand into occurrence and magnitude components while simultaneously leveraging shared information across related time series through embedding-based representations and lightweight temporal modeling. The framework is evaluated using an industrial automotive spare parts dataset characterized by high sparsity and short observation windows. Results show that the proposed approach achieves competitive forecasting performance compared to strong benchmark models, including LightGBM with a Poisson objective. Importantly, the findings reveal that explicitly modeling demand occurrence substantially improves accuracy in zero-demand periods, while alternative models tend to perform better in magnitude estimation during non-zero demand. These results highlight a fundamental trade-off in sparse demand forecasting between occurrence detection and magnitude estimation, suggesting that model selection should align with operational priorities. The proposed framework provides a practical and interpretable approach for short-horizon decision support in intermittent demand environments.
Discriminating hyperplastic polyps from sessile serrated adenomas (SSA) in colorectal histopathology is clinically consequential, with documented interpathologist agreement of only [Formula: see text]. Existing deep‑learning approaches treat hematoxylin and eosin (H&E) images as entangled RGB tensors, discarding biologically distinct stain signals. We propose Stain-Physics-Guided Representation Disentanglement (SPGRD), which recovers this information by decomposing H&E images into hematoxylin and eosin concentration maps via Beer–Lambert inversion, routing each through an independent DINOv2-S Vision Transformer encoder, and combining the three streams through a Transformer fusion module with learnable softmax pool weights. On the MHIST benchmark, SPGRD achieves accuracy of [Formula: see text] and AUC–ROC of [Formula: see text], an [Formula: see text] percentage point gain over a single-encoder baseline under identical protocol, using only [Formula: see text] trainable parameters ([Formula: see text] of total). The improvement is statistically significant across all seeds ([Formula: see text], Fisher-combined DeLong test) and Cohen’s [Formula: see text] of [Formula: see text] exceeds documented interpathologist agreement. Five stabilization techniques constrain per-seed accuracy SD to [Formula: see text] and AUC–ROC SD to [Formula: see text] across five independent training runs, providing the low-variance reproducibility required for clinical decision-support deployment on a small, imbalanced dataset. Head-only transfer to CRC-VAL-HE-7K yields AUC–ROC of [Formula: see text] with 770 parameters, confirming domain-general feature generalization across scanner and staining variability.
Such conditions as chronic obstructive pulmonary disease (COPD), pneumonia, and upper respiratory tract infections (URTI) form a significant health concern worldwide, which is why the development of effective diagnostic methods to detect them early deserves attention. The current study presents a deep learning model devoted to classifying audios of respiratory sounds – sourced from the ICBHI Respiratory Sound Database – into one of four classes: normal, crackles, wheezes, and combinations of crackles and wheezes. Preprocessing of the audio data involved resampling and segmentation, followed by the use of Short-Time Fourier Transform (STFT) to produce spectrograms. A set of data augmentation techniques, including time stretching, pitch shifting, and adding noise, was applied to expand the dataset. Unlike prior work, our study introduces a lightweight yet high-performing spectrogram-based CNN framework (ResNet18 and ResNet50) that is validated under patient-level cross-validation using the ICBHI dataset, addressing real-world deployability and generalization challenges in clinical settings. The imbalance between classes was handled using Focal Loss, which significantly improved the detection of minority class patterns such as combined crackles and wheezes. The ResNet18 and ResNet50 were trained with GroupKFold cross-validation to ensure strict patient separation. The ResNet50 achieved the best average harmonic mean score of 0.8429 and specificity of 0.9232.
Accurate long-horizon stress forecasting from multimodal data remains challenging due to non-stationary temporal dynamics, intersubject variability, and adaptive optimization mechanisms. This paper proposes an Adaptive Bio-Psychological Stress Prediction Framework with Intelligent Optimization (BioPsych-SPIO) for 24-hour stress progression forecasting in cardiac patients. This multimodal framework integrates Capsule Networks (CapsNet), Attention-based Bidirectional Long Short-Term Memory (BiLSTM), and Bidirectional Encoder Representations from Transformers (BERT) to jointly learn the patterns of Electrocardiogram (ECG) signals, Diabetic parameters, and textual sentiment data. This hybrid probabilistic stress transition model with Markov-based forecasting is used for early risk escalation prediction and alert scheduling. To enhance temporal learning, the Novel Cognitive-Aware Emotionally Adaptive Improved Harris Hawks Optimization (COEA-IHHO) strategy is introduced for dynamic hyperparameter tuning, modality-aware feature weights, and convergence regulation. Further, this framework is incorporated with an application-level visualization module that delivers patient-specific stress progression monitoring and threshold-based alert scheduling. The framework was tested using real-time data collected from 150 patients with cardiac abnormalities. The model achieves a cross-validation accuracy of 97% and an AUC of 0.9458. Independent testing on real-time clinical subject-wise data demonstrates consistent performance of 89% accuracy, indicating strong generalization capability. Benchmark testing confirms the optimization algorithm’s convergence stability, uncertainty robustness, and feature-selection capability.
Medical institutions produce large volumes of unstructured medical data that require the extraction of relevant medical information used for performing clinical studies and statistical analysis. This paper examines the application of large language models (LLMs) for named entity recognition (NER) in the medical domain, with a focus on their practical usefulness in clinical settings. We evaluate both prompt-based and fine-tuned approaches using the MACCROBAT2020 dataset, which includes clinical case reports annotated with biomedical entities. Furthermore, we extend our evaluation of the fine-tuning methodology to three biomedical NER datasets: QUAERO, NCBI, and E3C. The study compares the performance of several open-source LLMs against baseline models, using exact and relaxed F1 scores across multiple entity types. Fine-tuned LLMs achieved higher strict-match accuracy and produced more reliable structured outputs than prompt-based methods, GLiNER variants, and supervised BERT baselines, particularly for complex medical entities. However, BERT-based encoders remained substantially faster and competitive under relaxed matching. Their performance across multiple datasets, although uneven across languages and annotation schemas, together with efficient operation when quantized, indicates potential for clinical pilot studies. Our code is publicly available on GitHub (https://github.com/erikcalcina/llm-medical-ner) under the MIT license.
Traditional Enterprise Resource Planning systems impose complex menu-driven workflows requiring extensive training, limiting accessibility for non-technical users. This paper presents a five-pillar framework for agentic AI-driven ERP systems that transforms enterprise management from complexity to conversation. Integrating Large Language Models, Multi-Agent Systems, Natural Language Processing, Intelligent Process Automation, and Security Architecture, the framework enables natural language interfaces for intuitive enterprise interaction. Preliminary validation across three enterprise scenarios (finance, HR, supply chain) achieves 28 of 31 criteria passed (preliminary proof-of-concept, n = 3 scenarios). System dynamics analysis identifies four feedback loops governing adoption dynamics. Comprehensive empirical validation through enterprise deployments remains essential future work.
The Nurse Rostering Problem, also referred to as Staff Scheduling or Personnel Timetabling in the literature, is a combinatorial optimization problem of great practical importance in many fields (healthcare, manufacturing, etc.) that has been addressed using a wide variety of methods over time. Solving this problem is essential for building cost-effective and balanced timetables that satisfy a diverse set of real-world constraints. This paper offers a comprehensive and in-depth analysis of the recent advancements and future research directions in the nurse rostering field, covering works published between 2019 and 2025 inclusively. While examining a substantial number of contributions, the study provides a special focus on emerging machine‑learning (ML) approaches, highlighting their role and performance in addressing this problem, their strengths in overcoming the limitations of state-of-the-art non-ML methods, as well as open challenges and opportunities related to their development and integration. To the best of our knowledge, this is the first survey that captures the emergence of machine‑learning methods in the nurse rostering field.
Recent studies show that generative models can experience model collapse when they are repeatedly trained on or rely on their own outputs. Over time, the share of synthetic content increases relative to human-authored material, reinforcing existing patterns and reducing informational diversity. This dynamic is not confined to model training processes; it can also emerge organically within organizations as employees increasingly rely on AI assistants to generate documents, summaries, reports, and responses. In such environments, organizational knowledge systems may drift toward textual echo chambers, in which AI-generated content circulates, is repeatedly retrieved as contextual input, and subsequently amplifies itself with diminishing novelty. While echo chambers are well documented in political communication and social media, they have not yet been theorized or empirically examined in AI-mediated organizational knowledge systems. In this study, we first demonstrate - through controlled simulations across multiple real-world text domains - how recursive AI assistance gives rise to echo-chamber effects in organizational corpora. We then introduce a diagnostic framework that continuously monitors an organization's knowledge base and detects when it is drifting toward echo-chamber formation and collapse. This study contributes to the theory of AI-driven knowledge ecosystems and offers practical guidance for sustaining healthy, diverse, and resilient organizational information environments.
Customer churn prediction is essential for sustaining business performance and competitive advantage in the telecommunications industry. Early identification of churn enables service providers to design effective retention strategies. This study proposes a fine-tuned bagging-based ensemble deep neural network for telecom churn prediction. The proposed model is evaluated against several ensemble and standard machine learning classifiers using two imbalanced telecom datasets. Advanced oversampling and hybrid balancing techniques are applied to address class imbalance, and model performance is assessed using multiple evaluation metrics. Experimental results show that the ensemble DNN, particularly when combined with SMOTE-ENN, achieves the strongest overall and most balanced performance on both datasets compared to the evaluated approaches, demonstrating its robustness for telecom churn prediction.
Microclimate forecasting is critical to smart agriculture and climate-sensitive systems, particularly in greenhouse environments, where fine-grained predictions guide energy-efficient, adaptive controls. Raw sensor data suffer from high dimensionality, redundancy, and spurious correlations: challenges that traditional models cannot address without preprocessing and dimensionality reduction. Although feature selection (FS) using mutual information (MI) provides a solution, MI-based methods do not consider causality and therefore select noninformative variables. This study proposes the causal entropy swarm optimizer (CESO), a novel FS and forecasting framework that integrates conditional MI, causal inference, and a hybridized optimizer based on the gray wolf optimizer (GWO) and particle swarm optimization (PSO). By leveraging causality-aware temporal dependencies and MI-based selection criteria, CESO identifies the most informative subset of features from environmental time series data and significantly enhances prediction accuracy. Comparative experiments with baseline methods - PSO, the whale optimization algorithm, the firefly optimizer, and GWO - in terms of root mean squared error, mean absolute error, and coefficient of determination (R 2) confirmed the superior convergence speed and forecast accuracy of CESO. The proposed method provides a reproducible pipeline for microclimate intelligence systems by achieving an error reduction of up to 25% over standard metaheuristics while maintaining high interpretability and computational efficiency.
This study aims to develop a robust, multi-target algorithmic trading strategy for Bitcoin by leveraging advanced hybrid deep learning architectures. Our methodology first involves training and rigorously evaluating ten distinct models - including Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), and various attention-based encoder-decoders - on hourly data augmented with critical technical and on-chain metrics. Models were optimized across multi-target regression (forecasting metrics like pullback depth, and volatility regimes) and multi-target classification (predicting directional signs and breakout signals). Feature selection utilized Mutual Information (MI) and Permutation Importance (PI) to ensure model relevance. Models incorporating both LSTM and Attention mechanisms, particularly the Encoder-Decoder structure, demonstrated superior performance in the regression tasks, while the standard LSTM-Attention model led in the classification tasks. Finally, we translated these top-performing models into four distinct algorithmic trading strategies, integrating their outputs through Mamdani Fuzzy Inference Systems and Average True Range (ATR)-based risk management protocols. Walk-forward backtesting revealed that while the classification-based strategies achieved modest profitability, translating complex regression forecasts directly into profitable trading decisions remains challenging in highly volatile markets. The results underscore the power of Attention mechanisms in capturing market dynamics and confirm the utility of Fuzzy logic in synthesizing diverse ML outputs for actionable trading decisions.
Traditional approaches for predicting atomic energy levels, such as experimental measurements and computationally demanding calculations, are time-consuming and complex. The emergence of artificial intelligence (AI) provides a promising alternative. This study explores the application of AI algorithms to predict atomic energy levels, which are fundamental to understanding atomic structure. We first employed a neural network, one of the most widely used predictive techniques, and then advanced to the XGBoost algorithm. XGBoost effectively captures the complex relationships between quantum numbers and electronic configurations, providing more accurate and efficient predictions than the neural network. To construct the dataset required for training and evaluation, we computed the energy levels of the sodium-like chromium ion (Z = 24, Cr XIV) using two established atomic codes. The first is the pseudo-relativistic Hartree-Fock code with configuration interaction and relativistic corrections (Cowan's HFR method). The second is AUTOSTRUCTURE (AUTOS.), which incorporates Breit interactions and quantum electrodynamics contributions. In addition to energy levels, radiative lifetimes and Land & eacute; g-factors were also calculated. The predicted results were compared with previously reported data for validation. This research demonstrates the potential of AI-based methods in atomic data analysis and provides a foundation for further investigations in bridging atomic physics with artificial intelligence applications.
Class imbalance significantly hinders sentiment classification, as minority classes - despite their semantic importance - are underrepresented during training. Traditional undersampling methods often rely on random removal or local heuristics, which can eliminate informative samples and disrupt the data's global structure. To overcome these limitations, this study introduces two clustering-based undersampling methods, ClusUS and ClusUSCosine, designed to reduce redundancy in the majority class while preserving representative and meaningful instances. These approaches leverage structural and similarity-based selection to improve decision boundaries and enhance minority-class recognition. The proposed methods are evaluated on five Arabic and English benchmark datasets using two feature representations (TF-IDF and AraGPT2) and five deep learning models (CNN, LSTM, BiLSTM, RNN, and RNN-LSTM). Performance is measured using G-Mean, Macro F1, PR-AUC, and ranking-based stability analysis. Results show that traditional undersampling techniques are highly sensitive to feature representation, whereas ClusUS and ClusUSCosine deliver more consistent improvements. ClusUSCosine often achieves the best G-Mean, indicating strong minority-class performance, while ClusUS remains stable across settings. The study also highlights discrepancies between PR-AUC and balanced metrics, suggesting that ranking alone may overestimate performance. Overall, the proposed methods provide a robust, similarity-aware solution for imbalanced sentiment classification.
Semantic matching is foundational to intelligent scheduling, fault diagnosis, and procurement in Industrial Internet platforms. However, existing approaches lack domain robustness and interpretability when facing noisy, heterogeneous, and terminology-rich industrial texts. Large Language Models (LLMs) promise stronger semantic reasoning, yet direct deployment is hindered by massive parameter scales, high latency, and prohibitive resource demands, especially in edge environments. We address these gaps by leveraging high-capacity LLMs as teachers to build a reasoning-aware semantic matching framework tailored for the Industrial Internet. (i) We employ few-shot prompting with Chain-of-Thought (CoT) reasoning to couple accurate predictions with interpretable reasoning traces. (ii) We construct a reasoning-enriched dataset that augments each pair with CoT explanations, quantitatively vetted for fluency, coherence, diversity, and consistency. (iii) We introduce a knowledge distillation strategy that transfers teacher reasoning to compact student models, preserving interpretability while enabling efficient edge deployment. Extensive experiments show that the distilled compact model outperforms conventional neural baselines and even surpasses its teacher in accuracy, achieving an F1-score of 87.4% and Accuracy of 87.9%, while reducing inference latency from 9.5 s to 1.6 s and GPU memory from 2684 GB to 28 GB. These results highlight the pivotal role of high-capacity LLMs as reasoning teachers and demonstrate a practical path to trustworthy, interpretable, and edge-deployable semantic matching for the Industrial Internet.
Ground-level building damage assessment captures critical structural details that remain invisible in satellite imagery, yet this perspective remains severely underexplored in current research. We address this gap by introducing the first war damage dataset for ground-level building segmentation, comprising high-resolution side-view images of war-affected Ukrainian buildings with pixel-wise annotations across six semantic classes: Other, Building, Roof, Damage, Damaged Roof, and Broken Window. To preserve original image resolution while enabling efficient deep learning, we employ a patch-based strategy that divides each image into fixed-size regions, generating thousands of training samples from the original dataset. We propose an embedding-enhanced U-Net framework that enriches each patch with global ConvNeXt-Large embeddings and positional encodings to provide scene-level context and spatial awareness. We systematically evaluate different encoder architectures across 48 configurations, testing both simplified three-class and complex six-class segmentation tasks with and without embedding integration and Felzenszwalb superpixel post-processing. Results demonstrate substantial performance gains from embedding integration: ResNet-50 achieved +7.81 pp IoU improvement, reaching 0.7743 IoU and 0.8982 F1-score for three-class segmentation, while DINOv2 attained optimal six-class performance with 0.4711 IoU and 0.7462 F1-score, representing a + 4.65 pp IoU gain.
GenAI copilots for network operations rely on rapid retrieval of historical evidence for incident triage and root cause analysis (RCA). Yet retrieval is often evaluated solely on relevance, while tail latency - a first-class QoE constraint in operator-in-the-loop workflows - is rarely reported. We benchmark retrieval of log templates used as RCA evidence units across three Loghub datasets: Blue Gene/L (BGL), HDFS, and Thunderbird. We compare BM25, dense retrieval, hybrid reciprocal rank fusion (RRF), and contextual re-ranking with temporal recency and a proxy locality prior. Dense retrieval is the strongest overall quality-latency baseline and outperforms BM25 on quality across all three datasets. On BGL, Dense attains ${ m{MRR@10}} = 0.690$MRR@10=0.690, ${ m{nDCG@10}} = 0.760$nDCG@10=0.760, and ${ m{P95}} = 772$P95=772 ms; on Thunderbird, Dense again leads with ${ m{MAP}} = 0.170$MAP=0.170 and ${ m{Recall@10}} = 0.160$Recall@10=0.160. HDFS shows the strongest contextual gains because its anomalies are tightly time-localized at the block level, unlike the more dispersed BGL and Thunderbird settings; in those settings, LAR-Rerank reaches ${ m{MAP}} = 0.624$MAP=0.624 and ${ m{Recall@10}} = 0.496$Recall@10=0.496. These results show that dataset structure strongly shapes the value of temporal and locality priors, while Dense remains a robust default retrieval strategy. Code: https://github.com/Mustafa3946/rca-retrieval-bench.
