Accurate throughput estimation and control are critical for HAS over volatile mobile links. We present WISE, a client-side ABR that adapts to measured variability. WISE introduces the Weighted Harmonic-Exponential Mean (WHEM), which blends an EWMA with a weighted harmonic mean using a coefficient of variation gate: It leans toward a fast EWMA when links are calm and toward conservative WHM as burstiness grows. A variance-scaled buffer cushion further sizes the guard band to real-time dispersion and segment size, while a closedform QoE score balances bitrate, smoothness, and stall risk. All computations are constant-time and do not require offline training. On 5G traces, WISE improves mean quality by up to 46.1% and reduces switching magnitude by up to 53.3% versus state-of-the-art baselines, with competitive stall rates, yielding robust gains under both stable and highly volatile conditions.
This study analyzes the spatial characteristics of regional railway use in Japan by integrating large language model (LLM)-based classification of Instagram captions with geographic information. Posts mentioning 15 railway lines were categorized by an LLM (Gemma3-27B) into “tourism,” “daily,” and “other,” then linked to nearby stations and points of interest (POIs). Entropy and evenness were calculated to assess posting diversity and concentration. Results show that stable-demand lines exhibit balanced distributions of tourism and daily posts, latent-demand lines concentrate on specific tourist sites, and sparse-demand lines show dispersed but low posting activity. These differences reveal structural characteristics of regional communication and railway use. The findings demonstrate that combining LLM-based text understanding with spatial analysis can effectively visualize local interest structures and contribute to sustainable regional mobility planning.
Brain tumor segmentation is challenging due to the high heterogeneity in tumor size and shape. Standard U-Net architectures, relying on fixed $3 \times 3$ convolutional kernels, often struggle to capture features across varying scales effectively. In response to this challenge, we introduce CAMS-UNet, a novel architecture designed to integrate a dynamic multi-scale perception mechanism directly into the encoder. The proposed CAMS-Block integrates two pathways: 1) a Multi-Scale Path utilizing parallel convolutions ($1 \times 1$ to $7 \times 7$) to capture finegrained to coarse contextual details simultaneously, and 2) an Attention Path employing a Squeeze-and-Excitation mechanism with Softmax normalization to adaptively calibrate and select the most relevant receptive fields for each pixel. We evaluated CAMSUNet on the BraTS 2020 benchmark. Experimental results demonstrate that CAMS-UNet outperforms the baseline U-Net (DSC 0.5232) with a Dice Score of 0.5245 and achieves higher Sensitivity (0.5456), indicating an improved ability to detect tumor boundaries and minimize false negatives. These findings confirm that substituting fixed kernels with context-aware multiscale aggregation significantly enhances the model's robustness in handling heterogeneous anatomical structures.
Large language models (LLMs) enhanced with retrieval-augmented generation (RAG) and multimodal inputs are increasingly used as interfaces to organizational knowledge. However, their effectiveness in specialized, non-English enterprise settings-such as Thai HR support-remains largely unclear. In many Thai organizations, employees frequently ask detailed HR-related questions, but the relevant information is scattered across internal webpages, PDF manuals, announcements, and images, making it difficult for generic LLMs to provide accurate, policy-consistent responses. To address this issue, we develop a multimodal RAG pipeline that combines hybrid dense-sparse retrieval over a vector database and evaluate six LLM models on a private Thai Visual Question Answering (VQA) HR dataset consisting of 226 questions and reference images across five HR topics. The results show that recent multimodal models, especially Qwen2.5-VL, achieve the best performance, with the highest averages in correctness (0.54), relevance (0.75), and helpfulness (0.64), clearly outperforming older vision-language systems and a text-only reasoning model. For large-scale answer evaluation, we apply an LLM-as-a-judge approach using GPT-4.1 and Gemini 2.5 Flash. We found that it serves as a generally reliable, though imperfect, substitute for human evaluation.
This study presents a machine learning framework for predicting first-year undergraduate students at risk of low academic performance using data from 8,726 students (2013–2023) at a public university in Northeastern Thailand. A structured data preparation pipeline—comprising data cleaning, feature engineering, and feature selection—produced and selected the top 15 informative predictors. Five models (Logistic Regression, Decision Tree, Random Forest, XGBoost, and SVM) and a Stacking Ensemble were tuned using stratified 5-fold cross-validation. Logistic Regression achieved the best baseline performance (AUC = 0.8015; F1-at-risk = 0.5281). SMOTE-Tomek, evaluated experimentally, increased recall to 0.7282 but reduced precision, indicating limited suitability for operational use. The findings demonstrate that an interpretable logistic model provides effective early-risk identification to support timely intervention and enhance student retention in Thai higher education.
Converting tacit knowledge of physical movement to explicit knowledge is a novel challenge in the field of computermediated contexts. This study investigates the relationship between physical motion and onomatopoeia, focusing on the effect of verbal information on actual behavior. We analyzed the correlation between onomatopoeia interpretation and physical gripping motions in experiments using a custom pressure sensor and questionnaires. Data from six participants were measured and analyzed via hierarchical clustering. As a result, inter-pretation-movement alignment increased with repetition, and normalized mutual information rose by 5.21 % on average across participants. By contrast, the overall correlation between the stability of onomatopoeia interpretations and the reproducibility of physical movements was low (maximum absolute Spearman correlation 0.354), and individual differences were large (sample standard deviation of improvement 3.40 %). Therefore, models should learn interpretation and bodily expression separately and then integrate them. These findings show that onomatopoeia to movement conversion systems must account for individual differences and dynamic learning processes.
Recent advances in large-scale pre-trained language models have enabled dialogue systems to generate grammatically and semantically natural responses. However, existing chatbots often fail to produce emotionally appropriate and empathetic responses, resulting in mechanically perceived interactions. To address this issue, this paper proposes an emotion-controllable Japanese dialogue generation system that integrates emotion control directly into the generation process. The proposed system consists of a Japanese pre-trained language model (rinna Japanese GPT-2), a Plug and Play Language Model (PPLM)-based Bag-of-Words induction module, and a dialogue control interface that allows users to explicitly select emotion modes. By applying gradient-based induction to latent representations without retraining the base model, the system flexibly controls emotional nuances while preserving natural Japanese generation quality. Furthermore, emotions are treated as goal-oriented control elements that can be dynamically selected according to conversational purpose. This design enables more human-like, flexible, and empathetic dialogue compared to conventional sentiment-responsive chatbots.
This study proposes a model for analyzing Thai equities based on their sensitivities to global and local macroeconomic indicators. By using rolling correlations between stock returns—sampled across sectors according to FTSE Russell's Industry Classification Benchmark (ICB)—and selected global and local market indicators, we develop time-varying exposure profiles for each stock. Applying $K$ -means clustering to these profiles yields three behaviorally coherent groups. Principal Component Analysis (PCA) then reduces these exposure features into a two-dimensional regime space, with quadrants representing risk-on and risk-off conditions. Mapping the clustered stocks onto this regime space demonstrates how stock behaviors relate to prevailing macroeconomic conditions and aids transparent, data-driven investment decisions. Empirical results suggest that the proposed framework provides a clear and interpretable foundation for risk-aware investors, especially retirees seeking clarity in portfolio building.
Lane change maneuvers are among the most safety-critical actions for autonomous vehicles, as they require accurate perception, interaction prediction, and timely decision-making under dynamic traffic conditions. Unlike lane keeping or lane departure prevention, lane changes are intentional maneuvers that involve interaction with surrounding vehicles traveling in the same direction, particularly those approaching from the rear. This paper presents a simulation-based framework for evaluating Lane Change Assist (LCA) behavior using the CARLA simulator. The proposed system integrates sensor-fusion-based perception, short-term trajectory prediction, and safetyoriented decision logic to analyze overtaking maneuvers under multiple traffic scenarios. Simulation results illustrate the behavior of the proposed LCA framework in both base and failure scenarios, including cases where rear-vehicle awareness is absent. The simulation findings highlight the importance of dedicated lane change assistance mechanisms to support safe intentional lane change decision-making, particularly in scenarios where rear-side perception plays a critical role.
This study explores the application of machine learning techniques to support data-driven marketing for insect-based protein supplements in Thailand, using Brand $\mathrm{X}$ as the focal case. Survey data were analysed using three methodological approaches: classification, clustering, and association rule mining. To estimate consumers' likelihood of trying the product, three classification algorithms—Logistic Regression, Decision Tree, and Random Forest—were implemented, with Random Forest achieving the highest performance (76.11 % accuracy). K-means clustering identified two distinct consumer profiles: (1) premium- and variety-oriented consumers who emphasise product quality, certifications, and flavour diversity, and (2) price- and safety-conscious consumers who prefer familiar flavours and lower price ranges. FP-Growth and association rule mining further uncovered behavioural patterns, revealing that price concern is the most frequent and influential factor influencing purchasing decisions. The analysis showed that consumers motivated by weight control and those preferring vanilla flavour are often associated with the mid-tier price range $(801-1,200 \text{THB})$. Overall, the findings demonstrate that machine learning offers valuable insights for segmentation, product positioning, and pricing strategies within the emerging insect-based protein market, and can support more targeted and evidence-driven marketing decision-making.
Due to its unclear origins and lack of obvious early indicators, pancreatic cancer has a low early diagnostic rate. It results in a rather high death rate from pancreatic cancer. Imaging diagnosticians can diagnose pancreatic lesions more quickly, easily, and accurately by using artificial intelligence technologies to identify and diagnose malignant lesions. After analyzing the main image processing methods for pancreatic lesions from 2018 to 2025, we methodically evaluate the advantages and disadvantages of 2D & 3D deep learning models, transformer-based models, self-supervised pre-training models and hybrid feature fusion models. This study investigates fundamental methods for identifying and classifying pancreatic lesions, as well as specific surgical techniques, leading to the development of a FusionNet-based model for processing pancreatic cancer images. Multimodal pancreatic imaging data will be gathered in the future to create a better annotated database that will improve the robustness and generalization of the model.
Accurate trajectory prediction is essential for ensuring flight safety, autonomous navigation, and efficient control of unmanned aerial vehicles (UAVs). However, existing neural network approaches often focus on a single platform, limiting their generalizability across UAVs with different dynamics. This study evaluates three neural architectures for spatiotemporal trajectory prediction using high-frequency flight data. Two UAV configurations, such as a fixed-wing (Cessna) and a multirotor (Phantom), were simulated in XPlane 11 under identical environmental conditions. Flight data of latitude, longitude, and altitude sampled at 80 Hz were preprocessed using a sliding-window segmentation method to preserve temporal dependencies. Backpropagation Neural Network (BPNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) models were trained using standardised preprocessing and tuned hyperparameters. Results show that recurrent models outperform the feedforward BPNN in capturing temporal patterns. For the Cessna dataset, the RNN achieved balanced accuracy with a mean-squared error (MSE) of $5.5 \times 10^{-8}$, while the LSTM obtained the lowest MSE ($2.7 \times 10^{-8}$). In the more dynamic Phantom dataset, the BPNN remained competitive (MSE $\approx 1.8 \times 10^{-7}$), whereas RNN and LSTM maintained superior altitude prediction. These results highlight the suitability of lightweight recurrent architectures for real-time UAV trajectory prediction.
Many administrative tasks or work processes in organizations usually involve repetitive manual tasks that were time-consuming and prone to human error. To address these issues, the organization may decide to implement automation technology, specifically Robotic Process Automation (RPA). There are several RPA software tools to choose from with different features and usages. This paper aims to study and compare RPA software tools. The tools selected for comparative analysis are (1) UiPath (2) Power Automate (3) Automation Anywhere. The selection was based on market competition data from the widely recognized Gartner research. The comparative analysis includes execution time, work accuracy, variety of available actions, and complexity of development to achieve specific outcomes. The results of this comparative study can provide guidance to RPA tool selections that would ultimately reduce workload and time spent on manual tasks on work processes.
A graph-augmented, diffusion-aided framework for real-time fraud screening on Thailand's QR/PromptPay transactions is presented. Transactional attributes are first standardized with time-aware statistics, while missing fields are imputed via a conditional denoising diffusion model that provides feature-level uncertainty. Relational context among accounts, devices, phone numbers, and merchants is captured with a heterogeneous graph neural network; node embeddings are pooled over transaction endpoints and fused with imputed tabular signals and uncertainty summaries. A cost-sensitive classifier is calibrated (temperature + isotonic) and tuned under a Neyman-Pearson constraint to enforce a fixed false-positive rate. In offline, imbalanced evaluation reflecting production traffic, the method attains AUROC $=0.985$ and $\text{AUPRC}=0.842$, with $\text{TPR}=0.73$ at $\text{FPR}=0.1 {\%}$ and a normalized partial AUC over FPR $\leq 0.1 {\%}$ of 0.93. Calibration quality is evidenced by $\mathbf{E C E}=0.018$ and $\mathbf{N L L}=0.215$, and weekly alert volumes remain stable once the operating threshold is fixed. Ablations indicate that graph signals and diffusion-based imputation contribute complementary gains in the low-FPR regime. Limitations and deployment considerations (label latency, traffic scaling, temporal drift) are discussed, and avenues for continual, privacy-preserving learning and prospective A/B validation are outlined.
Accurate long-term forecasting of seasonal influenza is essential for public health preparedness. This study proposes a hybrid framework integrating cross-country lagged correlation analysis with an Extreme Gradient Boosting model to forecast influenza activity in the Republic of Korea. Weekly surveillance data from the World Health Organization FluNet system spanning January 2000 to April 2025 were analyzed to identify precursor countries whose influenza patterns exhibit strong positive correlations with those of Korea. Countries exceeding a correlation threshold of 0.6 and showing earlier epidemic peaks were retained as predictive features. Four experimental configurations were evaluated to examine how different temporal periods affect forecasting performance. Configuration 1 uses 2015 to 2018 data to forecast 2019, representing pre pandemic conditions. Configuration 2 uses 2022 to 2023 data to forecast 2024, reflecting post pandemic dynamics. Configuration 3 combines 2015 to 2019 and 2022 to 2023 data to forecast 2024 while excluding the disrupted years of 2020 to 2021. Configuration 4 uses continuous data from 2015 to 2023, including pandemic years, to forecast 2024. Among the first three configurations, Configuration 3 achieved the highest performance with an $R$ squared value of 0.708. Including the suppressed pandemic years increased performance further, with Configuration 4 achieving an $R$ squared value of 0.759. These findings demonstrate that integrating global surveillance signals with machine learning improves robustness and long-term forecasting accuracy for influenza activity in Korea.
This study presents a Vision Transformer (ViT-B/16)-based framework for facial emotion recognition (FER), offering significant improvements over traditional CNN-based approaches. A pilot study was conducted in a university classroom setting to evaluate the framework's performance in real-time e-learning environments. The system was integrated into an online learning platform, where it monitored students' emotional states during live lectures. Over a four-week period, the ViT-based FER model processed video feeds from 50 students, detecting emotions such as confusion, engagement, and boredom. By leveraging self-attention mechanisms, the proposed model captures global contextual relationships, achieving higher accuracy and robustness on FER2013 and AffectNet benchmark datasets used in emotion recognition research, particularly in challenging conditions like low light and occlusions.
Cyberattacks have become increasingly complex, exploiting vulnerabilities such as those listed in the Open Worldwide Application Security Project (OWASP) Top 10 to compromise systems and delay response times. Traditional incident response (IR) still depends heavily on manual procedures, which are prone to human error and slow mitigation. To address this challenge, this project proposes an AI-driven ChatOps (chat-based operations) Incident Response Bot that integrates automation, dynamic playbooks, and human-in-the-loop decision-making. The system operates in two main phases. In Phase 1 (Input Classification), the chatbot-built with Streamlit and LangChain-interacts with users, retrieves vulnerability data from the National Vulnerability Database (NVD) and Common Vulnerabilities and Exposures (CVE), and classifies incidents using a Gemini-powered large language model (LLM) with confidence evaluation. In Phase 2 (Playbook Generation), the bot automatically selects and merges YAML (YAML Ain't Markup Language)-based playbooks using NetworkX and applies organizational policies through the Open Policy Agent (OPA) before execution. This hybrid automation model enables rapid context-aware mitigation while preserving human oversight for high-impact actions. By integrating CVE intelligence, policy-as-code validation, and playbook orchestration, the framework establishes a scalable foundation for intelligent incident-response automation.
The main purpose of this research is to segment cleaning service providers of BeNeat Co., Ltd. to analyze behavioral characteristics across groups. Exploratory Data Analysis (EDA) and feature engineering were performed to construct features that represent provider performance. Clustering models were built using K-Means, Hierarchical (Agglomerative), and DBSCAN, with preprocessing through Robust Scaling and dimensionality reduction using Principal Component Analysis (PCA). Evaluation based on Silhouette Score, Davies-Bouldin Index, and Calinski-Harabasz Index showed that K-Means achieved the most distinct and effective segmentation, followed by Hierarchical Clustering, while DBSCAN produced less clear results. The K-Means analysis identified three segments: High Performers (high revenue and long service hours), Steady Providers (consistent and moderate), and Low/Inactive Providers (minimal activity).
This work uses ontology engineering and data mining to create and deploy a Semantic Web healthcare chatbot to enhance patient information support systems, which are becoming more crucial in digital healthcare. Healthcare organisations may use the Semantic Web to handle massive amounts of machinereadable and intelligible medical data. Semantically enhanced knowledge representations and automated analytic methods increase chatbot accuracy, relevance, and contextual comprehension, reducing rule-based systems' inconsistent outputs. This topic uses ontologies, linked data, and semantic web technologies to improve healthcare data interoperability and enable the accurate interchange of medical knowledge, records, clinical advice, and research. Over 64% of patients prefer automated conversational interfaces for preliminary medical guidance, and nearly 72% of healthcare organisations have adopted AI-enabled triage or support systems, indicating a growing societal reliance on automated health information platforms. Ontological analysis improves information comprehension. Hospital information management and patient self-service systems benefit from semantic chatbots' 89% retrieval accuracy and 43% reduction in irrelevant response production. An ontology is the foundation of any domain-specific knowledge representation system that removes data inconsistencies to assure validity.
Personalized support is increasingly needed in programming education, where students differ widely in prior knowledge, learning pace, and problem-solving behavior. This study presents a real classroom deployment of an adaptive learning system enhanced with behavioral analytics in an undergraduate programming course. The system recommends the next suitable task level using behavioral features such as lines of code, cyclomatic complexity, correctness, attempts, and time spent. In this course, 24 students voluntarily activated and used the adaptive feature, while 62 classmates who used the standard non-adaptive workflow served as the control group. Statistical analyses showed no significant difference in final coding exam scores, although adaptive users achieved slightly higher and more consistent outcomes. K-Means clustering revealed three behavioral profiles among active users: Fast Accurate, Consistent High, and Persistent Low, distinguished by engagement intensity and coding efficiency. These profiles were associated with meaningful differences in course grades. The results demonstrate the value of behavioral analytics for understanding learner diversity and informing more responsive adaptive programming support.
