In the context of rapid economic development and globalization, traditional cultural values such as Cai Luong, a Vietnamese theatrical art, face declining engagement, particularly among youth. This study addresses the challenge by first conducting a quantitative survey of domestic and international tourists in Ho Chi Minh City, which revealed strong demand for cultural experiences enhanced by digital technologies. To meet this need, we developed VCaiLuong, a cloud-based microservice system comprising Visitor and Admin applications. The platform integrates a virtual museum, live Watch Parties, and a digital archive, enabling interactive engagement with Cai Luong heritage. Features such as 3D asset management, real-time chat, notifications, bookings, and background job automation further enhance performance and scalability. By comparing existing platforms and identifying design gaps, this study demonstrates the role of digital technologies in heritage preservation. VCaiLuong offers a sustainable model for safeguarding and promoting Cai Luong through immersive and accessible cultural experiences.
In the context of sustainable digital transformation, healthcare is a security-critical domain that requires accurate and transparent medical image analysis. This paper presents Swin-FABNET, a cardiac MRI segmentation system that integrates a Swin Transformer backbone, UPerNet decoder, and a Fuzzy Attention (FAB) block for enhanced multi-scale feature learning. 3D cine-MRI volumes are preprocessed into 2.5D stacks with online data augmentation. Explainable AI (XAI) methods such as Grad-CAM, LIME, and SHAP are employed to visualize predictions and highlight clinically relevant cardiac regions. On the ACDC 2017 dataset, Swin-FABNET achieves an average Dice score of 0.86, outperforming the baseline and alternative designs. The results demonstrate the effectiveness of fuzzy attention in improving segmentation accuracy and contribute to the development of secure, transparent, and sustainable AI applications in healthcare.
Recognizing Vietnamese handwritten text poses significant challenges due to complex diacritics, tonal variations, and limited large-scale annotated datasets. Traditional deep learning-based OCR methods, such as CRNNs designed specifically for Vietnamese handwriting, have demonstrated strong accuracy and low word error rates when trained on targeted data. In contrast, Vision-based Large Language Models (Vision-LLMs) offer broader document understanding, multimodal perception, and downstream capabilities like question answering and content interpretation. However, these models often struggle with domain adaptation, linguistic nuances, and computational efficiency, especially in the context of Vietnamese script. This paper presents a systematic comparison of OCR-specific models and Vision-LLMs for Vietnamese handwriting recognition, examining model architecture, data dependencies, recognition performance, linguistic handling, flexibility, and deployment feasibility. The results indicate that while OCR-specialized models excel in accuracy and resource efficiency within narrow domains, Vision-LLMs provide greater adaptability across tasks - albeit with higher resource demands and sensitivity to script details. Finally, we propose future directions, including hybrid approaches, dataset expansion, and optimization strategies, to advance robust and versatile Vietnamese handwriting recognition systems.
Improving environmental sound classification is crucial for enhancing intelligent surveillance systems. In this study, we propose a hybrid approach that integrates data augmentation techniques with a deep learning architecture combining a Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), namely Hybrid Deep Learning–Data Augmentation (DLDA). The CNN (specialized in extracting local spatial features from spectrograms) captures discriminative patterns in the time–frequency domain, while the BLSTM (a recurrent neural network capable of learning temporal dependencies in both forward and backward directions) models long-range contextual information in audio sequences. To improve model generalization, we apply multiple data augmentation methods, including Gaussian noise addition, time shift, time stretching, pitch shifting, loudness scaling, and time delay. Experiments on the ESC-50 dataset demonstrate that our method achieves significantly higher classification accuracy compared to recent studies, confirming the effectiveness of combining advanced augmentation with our modeling.
Continuous, fine-grained water-quality monitoring is essential to prevent disease outbreaks and optimize yields in intensive shrimp aquaculture. We present an end-to-end IoT → ETL → AI framework that integrates 14 in-situ sensors, hourly cleansing and feature extraction, and sliding 10× 14 windows. Building on prior work, we adopt autoencoder-based anomaly detection trained only on normal data and contrast it with sequence forecasting, reflecting the trade-off often noted in aquaculture time-series literature between sensitivity and complexity. Our method instantiates three candidates: a 1D CNN Autoencoder, a Fully-Connected Sequence Autoencoder, and an LSTM Forecaster; anomaly thresholds are set from validation errors using μ +2σ . On a one-month dataset from an intensive shrimp farm, we apply stratified splits for train/validation/test. Results show the CNN–AE delivers the best precision–recall balance for early warning; FC–Seq–AE preserves high precision with reduced recall and smaller footprint; LSTM–Forecaster achieves zero false alarms but misses many events. The pipeline includes defense-in-depth security (mTLS-protected MQTT, AES-256 at rest, RBAC) and can support federated learning. Discussion highlights limitations (single site, early/mid-season coverage) and the need for multi-farm, late-season data and cost-sensitive learning to further improve recall and generalization.
This study proposes a fusion method that combines product metadata, user-item interaction data and visual cues using a Graph Neural Network (GNN) architecture for image recommendations. The study focuses on using latent features from a variety of modalities, such as product information, user behavior and photographs. We proposed an Attention Fusion method that enables customizable integration of GCN representations with textual and visual item features to improve recommendation accuracy. The system employs the Multi-View Graph Convolutional Network (MGCN) for image-based recommendations and compares its performance with popular models such as Multi-Modal Graph Convolutional Networks (MMGCN) and Bootstrap Latent Representations for Multi-modal Recommendation (BM3). The tests are conducted using three datasets: Clothing, Sports and Amazon Fashion. These datasets include latent features from text, images and interaction data. The results demonstrate that the proposed model outperforms baseline methods, particularly when sparse interactions are present.
This paper offers an enhanced technique to boost the efficiency of creating bird’s eye view (BEV) maps from 2D dash cam photos. We accomplish this by mapping items from 2D pictures to BEV space using the Detection Transformer network. In particular, we must map the relevant pixels of objects in 2D photographs to the polar coordinates on the BEV map image and ascertain the relative positions of items in those photos. Our approach generates semantic BEV mappings, covering both stationary and moving categories, by using an end-to-end learning transformer model that takes as input a monocular picture with an intrinsic matrix. The big dataset nuScenes, developed by nuTonomy and published in 2019 to aid in research on perception, autonomous driving, and traffic image prediction, is used to assess the approach. According to experimental data, the suggested approach greatly increases accuracy when compared to the state-of-the-art techniques now in use.
This paper presents UMIRA (Unified Memory-Integrated Retrieval and Task Alocation), a framework that integrates personalized memory-augmented retrieval with large language model (LLM)-driven task allocation and execution planning for elderly care in smart home environments. UMIRA combines an episodic memory module, which preserves and leverages user-specific interaction history, with a domain-specific Retrieval-Augmented Generation (RAG) pipeline to provide contextually grounded responses. In parallel, a multi-agent coordination module decomposes high-level user requests into a list of executable subtasks, allocates them across multiple embodied agents based on skill and capacity constraints, and manages their execution while preserving inter-task dependencies. The framework is evaluated in the AI2-THOR simulation platform using a curated elderly care knowledge base, with experiments covering retrieval accuracy, response generation quality, and multi-agent task execution efficiency. Experimental results demonstrate that UMIRA outperforms baseline methods, delivering higher retrieval precision, improved semantic alignment in generated responses, and more efficient, coordinated agent utilization. These findings highlight the potential of integrating personalized memory retrieval with LLM-based planning to enable adaptive, scalable, and user-centric assistance for aging-in-place scenarios.
In the context of an ever-fluctuating real estate market, accurate price forecasting is critical for investors and consumers. This study proposes a robust methodology for building and evaluating high-performance real estate price prediction models. The approach incorporates advanced data preprocessing, multi-dimensional feature engineering, and a hybrid modeling framework with optimal market segmentation. Through comprehensive statistical analysis including Elbow method, Silhouette score, Davies-Bouldin Index, and Calinski-Harabasz Index, we demonstrate that three price clusters (Low, Medium, High) provide optimal market segmentation with 85 ^2 > 0.99) indicating data leakage. After implementing rigorous validation protocols and removing 23 leaky features, LightGBM achieves the best performance with R ^2 = 0.843, demonstrating significant improvement over baseline methods while maintaining realistic performance within academic literature benchmarks.
This study presents a lightweight image retrieval system for a domain-specific dataset. The proposed system employs a Vision Transformer (ViT) with multi-head attention [1] to learn visual representations from a product image dataset collected from the Vietnamese e-commerce platforms. The model was trained entirely from scratch. The image feature vectors are stored and indexed in Qdrant to enable efficient similarity search. Experimental results indicate that the proposed approach provides improved clustering performance across several baseline models under our evaluation setting, while requiring fewer parameters.
Classification of intangible cultural heritage (ICH) images is a challenging task due to limited available image data and the high similarity between different classes, making them hard to distinguish. Consequently, traditional machine learning algorithms often achieve low accuracy on such tasks. Our investigation aims at proposing the Bagged CoAtNet (BagCoAtNet) to improve the classification of ICH images. Our BagCoAtNet algorithm learns an ensemble of CoAtNet models, where each CoAtNet is independently trained on a bootstrap sample (sampling with replacement) drawn from the original training image dataset. Each CoAtNet model is trained for a randomly chosen number of epochs, with a randomly selected number of top layers fine-tuned, with its final layer incorporating a random dropout rate. The prediction is determined by majority voting across all CoAtNet models in the ensemble. Empirical test results on the 17-category ICH image dataset demonstrates that our BagCoAtNet algorithm outperforms fine-tuned models like VGG16, DenseNet, CoAtNet, ViT and fusing deep network models. BagCoAtNet achieves a classification accuracy of 71.83
Pulmonary diseases pose significant health risks to humans, particularly with the emergence of COVID-19, which is characterized by rapid transmission and numerous complications. In situations where RT-PCR testing is limited, chest radiography serves as a crucial tool for rapid and effective diagnosis. The integration of artificial intelligence (AI) into healthcare has become increasingly essential, providing support to physicians in achieving accurate diagnoses and facilitating appropriate treatment strategies. This study proposes a deep learning-based approach for classifying pulmonary conditions using chest radiograph images into four categories: COVID-19, lung opacity, normal and viral pneumonia. A hybrid model is developed by combining three individual deep learning architectures: EffectiveNetB0, ResNet50, and MobileNetV2 - along with appropriate fine-tuning techniques to enhance adaptability to input data. Experimental results demonstrate that the hybrid model significantly outperforms each individual model in terms of accuracy, precision, recall, F1 score, and AUC-ROC. Specifically, the combined model achieves a precision greater than 95
Screening conversations often encode signs of psychological distress not only in lexical content but also in how utterances are delivered. We study whether large language models (LLMs) can better estimate depression risk when prosodic evidence is made explicit at prompt time. Using the DAIC-WOZ depression subset (AVEC-2017), we construct two prompting pipelines: (i) a transcript-only baseline and (ii) a multimodal variant that aligns 100 Hz COVAREP/formant features to each utterance and converts them into concise natural-language descriptors (e.g., low pitch, flat tone, long pauses). We keep the protocol constant across systems (official train/dev splits, PHQ-8 risk threshold at 10, and a JSON output schema) and evaluate on the development split. The multimodal prompting consistently improves Macro-F1 and PR-AUC by approximately 0.03 absolute over a strong transcript-only GPT-4 baseline, with parallel reductions in MAE and RMSE for PHQ-8 totals. Ablations show that verbalized acoustics are more effective than appending raw numeric feature summaries alone, and that preserving interviewer PHQ-8 prompts plus the immediate answers yields small but reliable gains. Our results suggest a practical path to multimodal sensitivity without training an audio encoder: align and verbalize prosody so that LLMs can reason over both what was said and how it was said.
Deepfakes threaten the trustworthiness of online media and public discourse. We present a lightweight yet effective hybrid for facial deepfake detection that couples MobileNet spatial encoders with sequence-level LSTM, probabilistic Hidden Markov Models (HMM) for temporal regularization, and Kolmogorov–Arnold Networks (KAN) for adaptive, interpretable nonlinear modeling. On FaceForensics++ and Celeb-DF v2—two widely used and challenging benchmarks—our best variant achieves strong cross-validated accuracy, outperforming both our baselines and previously reported methods on these datasets. We report precision, recall, F1-score, and AUC, and discuss training choices that yield robust generalization under video compression. Our results suggest that combining LSTM-based sequence modeling with HMM smoothing and KAN adapters is a practical path toward reliable, low-latency deepfake screening in real-world pipelines.
Community detection is a fundamental task in network analysis, but traditional modularity-based methods often overlook node attributes that provide crucial semantic information. We propose an attribute-aware null model that extends the Louvain algorithm by incorporating node feature similarity into the expected edge weight calculation. Our approach introduces a theoretically grounded modification that preserves modularity properties while capturing both structural and semantic coherence. Unlike existing attribute-aware methods that require deep learning or complex optimization, our method maintains the simplicity and scalability of Louvain while achieving improved performance. Experiments on Cora and Citeseer citation networks demonstrate that our method achieves competitive modularity scores and significantly better alignment with ground-truth communities compared to both classical and recent baselines. The proposed approach offers a practical solution for networks where node attributes provide meaningful community indicators.
Falls are a serious and common health issue among the elderly, especially in the context of the rapidly aging population. The consequences of falls extend beyond physical injuries such as bone fractures or cranial trauma; they also lead to reduced mobility, loss of self-care ability, and increased dependence on healthcare systems. In addition, prolonged psychological effects such as anxiety, decreased confidence in movement, and the risk of depression often occur in individuals who have experienced falls. Timely fall detection plays a crucial role in enabling early intervention and minimizing both physical and psychological complications. This paper focuses on the research and development of a mobile application system capable of detecting fall events, analyzing the level of medical urgency, and transmitting the victim’s status information to family members or relevant emergency services. The system is designed to shorten response time, thereby improving timely access to medical care and increasing the chances of recovery for elderly individuals involved in fall accidents.
With the growing reliance on mobile communication, Short Message Service (SMS) spam has become a persistent issue, disrupting user experiences and introducing security threats such as phishing attacks and fraudulent schemes. In Vietnam, SMS spam, which frequently appears in both Vietnamese and English, exerts a negative impact on a wide range of users. To address this pervasive problem, a bilingual spam classification approach is crucial. In this paper, we propose a hybrid model, VIESpam - Vietnamese-English Spam detection- that leverages the langdetect library in Python to identify English and Vietnamese messages, followed by the application of pre-trained language models, whose embeddings are processed by a BiLSTM network for final classification. Specifically, in terms of applied pre-trained language models, to enhance adaptability to bilingual contexts, we use DistilBERT and a reduced-layer PhoBERT for English and Vietnamese messages, respectively. Experiments on a 24,086-message bilingual dataset demonstrate that VIESpam achieves competitive performance. In particular, compared to traditional deep learning models, it delivers a high accuracy of 96.75
The advent of Generative Artificial Intelligence (GenAI) presents unique opportunities and challenges for higher education in developing countries, where the empirical research landscape is an emerging but scattered body of literature. This paper conducts a systematic literature review following PRISMA guidelines to consolidate this field, analyzing 33 studies from Scopus and Web of Science. Using a hybrid synthesis of thematic analysis and ADO-TCM framework, our findings reveal a research landscape dominated by qualitative, cross-sectional studies from a few nations. While scholars have focused on adoption factors and pedagogical roles, our most critical finding is a significant “Outcome Gap”: the literature overwhelmingly investigates adoption antecedents and intentions, while the actual use behavior of faculty and its tangible impact on teaching or research productivity remain sparsely explored. This review synthesizes these findings into an evidence-based gap map and proposes an actionable agenda to guide future research toward more methodologically diverse, context-aware, and impact-oriented inquiry in developing countries.
Accurate stock forecasts are most useful when paired with reliable uncertainty estimates for risk-aware decisions. We propose BiMamba Bayesian MAGAC, a unified architecture that couples a bidirectional selective state-space temporal encoder with a Bayesian graph module for cross-asset dependencies. Our contributions are threefold. (i) Temporal modeling: a bidirectional selective SSM (BiMamba) that captures forward/backward dependencies with linear complexity 𝒪(L) while preserving causality over the most recent L steps. (ii) Graph reasoning: a Bayesian Multi-head Adaptive Graph Attention Convolution (MAGAC) with: adaptive adjacency blending between a Gaussian locality prior and data-driven attention; node-conditioned, factorized spectral filtering that separates who interacts from how far information propagates; and DropEdge regularization. (iii) Uncertainty quantification: closed-form variance propagation under a diagonal approximation for 𝒪(N) cost, complemented by lightweight Monte Carlo via MC-Dropout for calibrated posteriors. Compared with methods using fixed adjacencies or expensive full-covariance estimates, our design yields efficient, uncertainty-aware predictions with a modest ∼ 2× overhead. Training employs a heteroscedastic Gaussian negative log-likelihood to jointly optimize accuracy and calibration end-to-end. Across benchmarks, the model delivers consistent predictive gains and better-calibrated risk estimates, supporting practical uses such as position sizing and downside protection. Implementation: https://github.com/ngngsonan/MAMBA_BGNN .
Building a large natural language inference dataset manually is a time-consuming task. The natural language inference dataset is composed of three classes: entailment, contradiction, and neutral. When building a Vietnamese NLI dataset, entailment samples may be collected from Vietnamese news web pages to reduce annotation costs. If the contradictions can be automatically generated, the cost will be reduced more. Therefore, a process has been proposed for building a Vietnamese NLI dataset. This process employed a high-quality large language model to generate contradiction samples. To verify if this process significantly helps, the TRAIN set and the TEST set were built using this process. The experiments showed that a large language model was useful to build a Vietnamese NLI dataset by manually filtering incorrect sentences. When using a large language model without manually filtering incorrect sentences, the quality of the dataset is low.
