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.
