Research & Theses

Plant diseases substantially challenge agricultural productivity and global food security. Hence, better intelligent and interpretable diagnostic frameworks are needed. An auto mated disease identification system can reduce the human effort in checking large farms, and early detection and identification will minimize the loss, which ultimately positively affects the economy. Traditional image-based deep learning models, particularly Convo lutional Neural Networks (CNNs), often struggle to distinguish visually similar diseases due to the absence of contextual information. To address these limitations, we present an innovative multi-modal deep learning framework that effectively combines visual and textual data to improve plant disease classification and segmentation. Initially, the framework incorporates a linguistically enriched Text Encoder, where disease-related descriptions are preprocessed using natural language processing (NLP) techniques to extract salient noun, numerical, adjective, and adverbial features. These refined textual representations are then encoded using a fine-tuned transformer-based language model, capturing domain-specific semantics crucial for disease differentiation. Concurrently, CNN-based Vision Encoder extract discriminative hierarchical features, which are dy namically fused with textual representations via a multi-head attention mechanism, en suring adaptive cross-modal feature alignment. Unlike conventional fusion techniques, our approach learns complex inter-dependencies between textual cues and visual pat terns, enhancing classification accuracy and segmentation precision. Finally, we demon strate our proposed framework’seffectiveness byevaluating it onthePlantDiseaseDiag nosis Multimodal Dataset (PDDM) andachieving state-of-the-art (SOTA) segmentation and classification performance.

The rapid advancement of Retrieval-Augmented Generation (RAG) models has transformed information retrieval and generation by integrating external knowledge with powerful generative capabilities. Despite these benefits, the environmental impact of deploying such systems particularly in terms of energy consumption and carbon emis sions remains an important concern. This study investigates the development and assessment of energy-efficient RAG archi tectures, focusing on two major approaches: Vector RAG and Knowledge Graph-based Retrieval System. Using academic content extracted from the Asian Institute of Technology website, both models were implemented to support question answering and contex tual information retrieval. A central component of the research involved measuring and analyzing the carbon footprint of each RAG architecture through detailed monitoring of CPUand GPUenergy usage.The comparative evaluation reveals differences in retrieval performance, accuracy, and energy demands, and identifies the most energy-intensive stages within each pipeline. Additionally, the study introduces optimization strategies aimed at improving computational efficiency while maintaining retrieval quality and response consistency. The findings contribute meaningful insights toward the design of sustainable AI systems that reduce environmental impact without sacrificing performance, supporting broader goals for environmentally responsible AI deployment.

Nepal has increasingly positioned e-government systems at the center of its public sector modernization efforts, yet significant gaps remain between policy formulation and on-the-ground implementation. This study examines how effectively national cybersecurity policies, such as the ICT Policy 2015, Cybersecurity Bylaws 2077, and the Digital Nepal Framework, are being applied across public agencies, the types of cyber threats these institutions commonly face, and how Nepal’s cybersecurity posture compares with Sri Lanka and Estonia. A mixed-methods design was used, combining a nationwide quantitative survey with in-depth interviews with officials from MoCIT, NITC, NCSC, NTA, and sectoral agencies. Descriptive statistics, reliability tests, non parametric analyses, and correlation matrices were triangulated with thematic coding to produce an integrated assessment. Findings show that policy implementation remains uneven, with low perceived threat detection despite recurring incidents such as phishing, ransomware, website defacement, and system downtime noted in interviews. Key implementation bottlenecks include limited budgeting, weak inter-agency coordination, dependence on external vendors, inadequate technical expertise, and the absence of standardized security audits. Comparatively, Sri Lanka demonstrates stronger institutional preparedness, while Estonia highlights the architectural and governance features of a highly mature ecosystem.Based on these insights, the study recommends strengthening budgeting and resource allocation, institutionalizing routine security audits, improving coordination mechanisms, expanding technical capacity, and reducing over-reliance on external vendors. The research contributes an empirically grounded E-Government Cybersecurity Implementation Gap Model and offers one of the most detailed evidence based assessments of Nepal’s e-government cybersecurity landscape to date.

Chest X-ray lung segmentation is essential for automated diagnosis of respiratory diseases, yet current approaches using either CNNs or Vision Transformers independently often is not fully captured all context of chest X-ray images . These images require both precise local boundary detection and global anatomical understanding due to overlapping structures and projection artifacts. While hybrid CNN-Transformer models have succeeded in computer vision, their application to medical image segmentation remains largely unexplored. This thesis presents hybrid architecture combining CNN and Transformer approaches for chest X-ray lung segmentation. The model employs parallel processing pathways where a CNN branch extracts local features and a Transformer branch captures global context. The key is a cross-modal attention mechanism enabling information exchange between branches. Through multi-head cross-attention, the CNN learns global context from the Transformer while the Transformer acquires spatial details from the CNN, creating enhanced representations that leverage both local precision and global coherence. The fused features are processed through a decoder with skip connections. Evaluation on standard datasets demonstrates improved segmentation performance compared to existing methods while maintaining computational efficiency. This research contributes the a hybrid CNN-Transformer architecture specifically designed for chest X-ray lung segmentation, establishing a foundation for future medical hybrid models.

Customer retention is a critical determinant of profitability in ride-hailing platforms, as the cost of acquiring new riders consistently exceeds the cost of retaining existing ones. This study explores customer segmentation and churn prediction using a large-scale dataset of ride-hailing trip records, incorporating behavioral signals such as trip frequency, spending patterns, service mode usage, and recency of travel. Motivated by recent mobility-focused research (Loureiro et al., 2025; Forecasting Client Retention, 2020; Comparative Analysis of Ride-Hailing vs Taxi, 2024), this work extends churn analytics beyond traditional dashboards by modelling disengagement using machine learning. User-level features were engineered to capture temporal engagement, cross-service usage, monetary contribution, cancellation behavior, and activity gaps. Churn was defined as prolonged inactivity, and predictive models—Logistic Regression, Random Forest, and Support Vector Machine—were trained and evaluated against a baseline majority classifier. Performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC to measure both discrimination ability and reliability under class imbalance.Findings indicate that customer behavior patterns—particularly ride frequency, inter-trip gap duration, and service type preference—are strongly predictive of churn. The results demonstrate that churn modelling can move beyond descriptive reporting into proactive churn risk identification, offering direct managerial value for retention planning. This study contributes to the literature by applying machine-learning-based churn prediction to the ride hailing domain and presents a data-driven foundation for segmentation-specific retention strategy design.

Telecommunication service providers face increasing challenges in managing the large volume of alarms generated by complex broadband networks. In particular, VNPT’s IP Broadband Network requires an efficient mechanism to detect, classify, and resolve alarms in real time to maintain service quality and operational efficiency. This thesis presents the design and development of a real-time alarm management dashboard that enhances fault management processes through improved visibility, automation readiness, and performance monitoring.The proposed system integrates data from three primary sources: the alarm dataset, the network device inventory, and port traffic statistics. Using a rule-based classification framework, alarms are categorized into either automatable or manual, enabling repetitive low-risk faults to be resolved through predefined actions, while critical issues are escalated to operators. The dashboard provides role-based views: operators receive real-time alarm notifications and automation options, while managers access KPI-driven insights. Key performance indicators (KPIs) such as Mean Time to Acknowledge (MTTA), Mean Time to Resolve (MTTR), and Automation Ratio are continuously calculated and displayed, allowing the system to measure and improve operational performance.The research adopts a design science methodology, progressing through problem identification, system design, implementation, evaluation, and refinement.The evaluation demonstrates that the dashboard significantly reduces MTTA and MTTR while increasing the proportion of alarms resolved automatically, thereby improving efficiency and service reliability. By aligning with industry standards (e.g., ITU-T M.3400) and incorporating ethical and security considerations, the proposed dashboard offers a practical, secure, and scalable solution to alarm management challenges in large-scale broadband networks.

This research addresses the limitations of the current Productivity and Quality Measurement System (NSCL) at VNPT Hanoi’s Information Operation Center (IOC), which relies on a legacy SQL-based database. The existing system suffers from rigid data structures, oversized and inconsistent job catalogs, lack of task segmentation, manual data entry for high-frequency operations, and the absence of real-time dashboards. These weaknesses result in unfair employee evaluations, operational inefficiencies, and limited decision-making support.The study aims to design and implement an improved, task-segmented evaluation framework that ensures fairness, consistency, and transparency. Using 2024 operational records from six departments, the research employs a mixed-methods approach: quantitative data analysis, qualitative interviews, and benchmarking against international standards such as ISO 9001, Lean IT, and TM Forum eTOM. The proposed system integrates MongoDB for flexible data modeling, task segmentation for fair scoring, and Business Intelligence dashboards for real- time KPI monitoring.Expected contributions include a standardized job catalog, a scalable database architecture, automation of repetitive tasks, and a unified KPI evaluation model that combines organizational objectives with task-based performance. The framework also establishes data governance mechanisms to enhance data integrity and proposes predictive analytics as a future extension to identify low performance early.

Artificial Intelligence has spread like a wildfire across the humancivilizationat this time. However, with the pervasive use of AI in almost every domains of life, there has been a significant level of concerns regarding the ethical aspects of it. Due to this reason, today several researchers have started working on how this technology can be ethically reasonable.Several frameworks and guidelines have been proposed by several researchers in this growing field. However, there is a lack of a framework or study that goes beyond the theoretical aspect and provides more real time and quantitative dimension to the ethical AI domain. This research addresses that gap by proposing a practical ethical compliance framework that introduces quantitative metrics across the AI lifecycle. Centered on four key ethical principles: Fairness, Accountability, Transparency, and Privacy, the framework consists of 39 metrics mapped across six AI lifecycle phases: data ingestion, data processing, model training, evaluation, deployment, and post-deployment. These metrics are designed to help AI Developers and Compliance Officers evaluate the degree of ethical adherence in AI systems in a measurable, repeatable manner. To validate the framework, an expert review process was conducted involving both technical and healthcare professionals. Experts rated the framework on clarity, validity, and practicality. Feedback was analyzed using descriptive statistics and thematic coding to extract key insights and improvement areas. This study offers a foundational step toward operationalizing ethical AI in healthcare through structured, lifecycle-aware evaluation.

The increasing intricacy and prevalence of cyber threats in modern computer networks highlight the need for effective anomaly detection systems to protect sensitive information. Traditional methods face challenges such as limited real-time processing, reliance on simple binary classification, and inadequate evaluation using realistic datasets. To address these issues, this research proposes a hybrid machine learning framework for anomaly detection. In the first stage, an autoencoder is used to learn latent represen tations of normal traffic, while an Isolation Forest algorithm detects anomalies based on anomaly scores. The Receiver Operating Characteristic (ROC) curve and Youden’s Index are employed to determine thresholds, which are then validated against the test labels of the UNSW-NB15 dataset to obtain baseline performance metrics. In the sec ond stage, supervised models including Decision Tree, XGBoost, and Random Forest are trained on the latent features, reconstruction error of the autoencoder, and anomaly scores from the Isolation Forest. Among these, Random Forest achieved the best per formance, significantly improving upon the unsupervised baseline, with an accuracy of 98.81%, precision of 92.00%, recall of 99.25%, F1-score of 95.49%, and a false posi tive rate of only 1.25%. To enable real-time usage, the framework is deployed with the Elastic Stack (ELK), allowing automated alerting, continuous monitoring, and visual ization of security events. The deployed system is further evaluated using real-world NetFlow and Syslog data collected from the Operational Technology (OT) network of the Provincial Electricity Authority (PEA). Controlled attack scenarios including TCP, UDP, and ICMP flood attacks are generated using the Nping tool to validate real-time anomaly detection. The results confirm that the proposed framework performs effec tively under realistic operational conditions and is suitable for practical deployment in critical infrastructure environments.

The increasing adoption and scale of automated machine learning (AutoML) systems have raised concerns regarding their environmental impact, particularly due to the substantial energy consumption and associated carbon emissions incurred during model training. While AutoML frameworks simplify model development, they often intensify resource utilization through extensive hyperparameter tuning and search processes. To address these challenges, this study proposes GreenMLX, a modular energy-efficient AutoML pipeline that integrates multi-objective optimization using Optuna, enabling simultaneous optimization of predictive accuracy, energy consumption, and training time. The pipeline also incorporates early stopping strategies, real-time emissions tracking via CodeCarbon, and support for model compression techniques aimed at reducing the environmental footprint. GreenMLX was evaluated against baseline AutoML frameworks—FLAML and AutoGluon—across ten benchmark tabular datasets from OpenML. Experimental results demonstrate that the Green pipelines achieved substantial reductions in energy consumption (up to 99.88%) and carbon emissions (up to 99.94%), with minimal degradation in model accuracy (within 3% of baseline performance in most cases). A detailed case study on the Household Power Consumption dataset further confirmed that Green pipelines effectively reduce energy usage, carbon emissions, and training time without compromising predictive performance.Moreover, explainability analyses using SHAP (SHapley Additive exPlanations) provided in sights into model behavior, validating that GreenMLX models maintained interpretabil ity while achieving sustainability objectives. These findings highlight the feasibility of integrating energy-efficient practices into AutoML workflows, offering a practical path toward environmentally sustainable machine learning without sacrificing model effectiveness or transparency.

Generative AI tools such as GPT are becoming increasingly common in programming education due to their potential to enhance student learning. However, these tools fre quently generate incorrect or misleading information, a problem known as ’hallucina tion’, which can negatively affect learning experiences. This research aims to address this issue by developing and testing GPT-based tutors designed to be more reliable and accurate. Specifically, the study proposes comparing two augmentation techniques: Retrieval-Augmented Generation (RAG) and Document-Augmented Generation. While retrieval-based AI support has been studied in general educational settings, its applica tion within GPT-powered tutoring for programming, particularly on difficult-level pro gramming tasks, remains underexplored. Experiments involve programming assign ment problems sourced from the Machine Learning course taught at the Asian Institute of Technology. Three GPT-based tutors are developed: two Document-Augmented GPT tutor utilizing provided course materials (one using Custom ChatGPT from OpenAI and one using LangChain), and a RAG-Augmented GPT tutor. The three developed tutors are then compared with two non-augmented models: baseline ChatGPT and baseline ChatGPT in Study Mode. The study contributes by testing RAG in a domain where hal lucination risk is high and factual accuracy grounding is critical. The findings of this study will provide valuable insights and practical recommendations for educators look ing to integrate generative AI tools safely and effectively into programming education.

In this research, literature reviews generated using our ChatGPT and Gemini against human-authored reviews from three different academic domains were analyzed. Using a corpus of 180 articles, a mixed-method design that combines qualitative expert assessment (accuracy, clarity, completeness, and source links) with quantitative measures (ROUGE scores, Cosine Similarity, sentiment analysis, and feature extraction with RoBERTa) has been designed. Both unsupervised clustering and supervised machine learning models were used. Such results suggest that the semantically closest paraphrased outputs from ChatGPT to human-authored reviews closely match the original reviews, where Gemini performs moderately lower but still acceptable, and full AI-generated reviews also show evidence of underdeveloped critical reasoning, referencing, and domain-specific depth, especially in narrative-heavy domains. The research demonstrates the promise and moral peril of using generative AI for scholarly work, recommending a hybrid model that combines human and machine intelligence to maintain the rigor, trustworthiness, and ethical use of such technology.

This final internship report presents an end-to-end customer experience (CX) transformation study for the True App, focusing on identifying critical service gaps and converting them into actionable Feature Requirement Documents (FRDs) to support the organization’s digital self-service roadmap. The project scope focuses on two core components: (1) a comprehensive gap analysis based on customer pain points, and (2) the development of a 10-FRD portfolio addressing the highest-impact improvement opportunities. The study adopted a mixed-method qualitative approach, including 12 intercept interviews across three age groups and gap scoring across seven high-volume service categories. Key findings reveal that digital literacy significantly influences adoption, with older users showing low awareness, low trust, and reliance on physical channels. Gap analysis identified five persistent issues across the True App: functional incompleteness, limited feature awareness, inconsistent policies between channels, complex navigation, and insufficient data tracking. To address these gaps, the report delivers a structured Gap-to-FRD framework, producing 10 prioritized FRDs covering network status transparency, diagnostic tools, unified package flows, simple add-on flows, device troubleshooting, eSIM activation, and complaint automation. These FRDs collectively serve as the foundation for a scalable CX transformation roadmap and offer design principles for enhancing usability, trust, and operational efficiency. The report concludes with recommendations for the True App’s future roadmap, including guided onboarding, real-time visibility features, simplified journeys, and a CX analytics dashboard for continuous monitoring.

With the rapid advancement of technology, AI-powered surveillance systems are becoming increasingly essential for security and monitoring applications. The AS Box, an AI-driven computer vision device, is designed to perform license plate recognition, motion detection, and other real-time surveillance tasks. However, configuring the AS Box for network connectivity remains a challenge, requiring users to manually connect peripherals and input Wi-Fi credentials, making the setup process complex and inefficient. This study focuses on the development of a mobile application to streamline the network configuration process of the AS Box. The application, developed using Flutter, facilitates hotspot-based pairing, Wi-Fi/LAN connectivity, and automated IP address retrieval. A FastAPI backend is integrated into the AS Box to handle network requests and ensure seamless communication between the device and the mobile application. The system also incorporates SQLite for local storage, allowing users to manage multiple AS Boxes efficiently. The development process follows the Agile methodology to ensure rapid iterations, software quality, and risk reduction. Extensive unit testing and quality assurance are carried out, focusing on backend stability, API reliability, and overall system performance. The final product undergoes rigorous testing phases, from internal validation to beta deployment, ensuring a robust and user-friendly experience before full-scale implementation.By automating network setup and device management, this mobile application enhances the usability and scalability of the AS Box, making AI-powered surveillance solutions more accessible, efficient, and user-friendly.

Precision agriculture supported by cutting-edge technologies, has emerged as a revolutionizing concept to optimize farming processes. The study examines the application of machine learning and deep learning techniques within the precision farming platform, focusing on crop and pesticide recommendation models. Previous studies have already demonstrated the beneficial potential of such technology in helping farmers make more informed decisions. However, one area of concern has been largely omitted in this context: uncertainty quantification in machine learning models. Indeed, real-world agriculture systems are uncertain or not easily predictable due to unpredictable environmental conditions and the complexity of interactions between numerous factors. The proposed study addresses this gap by introducing uncertainty quantification into crop and pesticide recommendation models, aiming to enhance robustness and reliability. The research not only seeks to improve the state-of-the-art results in experimented datasets but also contributes to the development of adaptive and resilient agricultural practices, fostering sustainable and efficient farming.The main objectives include the development of ML and DL models, incorporation of uncertainty quantification techniques, and an extensive evaluation of their performance, ultimately advancing agricultural decision-making for a more resilient and sustainable future. The experimental results demonstrate the superior performance of our ensemble model for crop recommendation, achieving an accuracy of 99.54%, surpassing existing studies. Furthermore, our developed Resnet152 model surpassed previous pest detection models, attaining an accuracy of 99.06%. Additionally, this study delves into the significance of uncertainty in ML/DL models through various ablation studies. A web application is also developed to demonstrate the usage of the proposed recommendation models.

This thesis assesses the impact of e-governmen ton public service delivery in Kathmandu Metropolitan City, with a particular focus on the Electronic Building Permit System (E-BPS). Thestudy utilizes a mixed-methods approach, combining quantitative data from a structured survey of 113 participants to assess perceptions of access, quality, efficiency, andparticipation levels in e-government services, with qualitative insights from in-depth interviews with 14 key stakeholders involved with the E-BPS. The first part of results re veal that enhanced access and quality of e-government services significantly boost user satisfaction and trust. Meanwhile, efficiency influences user willingness to engage regu larly, although it does not necessarily lead to deeper engagement. The second part of this study employs thematic analysis to illustrate how the E-BPS has improved efficiency and transparency, contributing positively to user satisfaction and system trust, despite chal lenges like digital literacy and system integration issues. The findings underscore the potential of e-government initiatives to transform public service delivery. However, they also highlight persistent barriers that could hinder their effectiveness. Addressing these challenges could further optimize the benefits of digital governance, providing crucial insights for policymakers engaged in digital transformations in developing countries. This research contributes to the broader discourse on implementing e-government so lutions and offers actionable recommendations for enhancing public service delivery through technological innovations

Automated product recognition plays a crucial role in the retail sector due to its myriad applications in Computer Vision. Object recognition technology enables retailers to efficiently monitor product placement and presentation on shelves. By analyzing images or video feeds of shelves, retailers can evaluate product arrangement, shelf compliance, and promotional displays. This data empowers retailers to optimize shelf layouts, ensure proper product positioning, and implement effective merchandising strategies to enhance customer engagement and boost profitability. This thesis focuses on utilizing the power of computer vision for product recognition on shelves and identifying empty spaces within visicoolers, leveraging a custom dataset for model training. The objective is to compare different models to determine their accuracy, efficiency, and suitability for product recognition tasks in retail environments.

Biodiversity supports ecological stability and human well-being by providing essential services like food security and climate regulation. Global initiatives emphasize conser vation, while ecotourism promotes responsible travel that sustains biodiversity and ben efits local communities. High Conservation Value (HCV) analysis further identifies key ecological, social, and cultural areas critical for conservation and tourism. Knowledge Graphs offer advanced insights by integrating complex ecological and touristic data.This study addresses the gap in understanding the biodiversity-ecotourism relationship by constructing an ontology-based knowledge graph for the Ranong Biosphere Reserve, Thailand. Semi-structured data from NECTEC, was processed in Protégé, revealing Biodiversity (HCV1), Livelihood (HCV5), and Cultural (HCV6) hotspots, along with novel insights from the integrated ontology. A user-friendly interface was developed to enable stakeholders to access data for informed decisions. The BioEcoThai ontology aligns with SDGs 12.b.1, 15.5.1, and 15.9.1 by supporting habitat conservation, biodiversity integration in planning, and sustainable tourism. It empowers stakeholders to identify High Conservation Value Areas, optimize resource use, and foster sustainable tourism and conservation efforts.

The complexity of crop production drives farmers to utilize a Farm Management Information System (FMIS) to manage their farms more effectively and efficiently. However, the existing FMIS are prohibitively expensive, particularly for smallholder farmers, the majority of farmers in the world. Providing an FMIS application that conforms to the smallholder farmers’ needs is a very tough task. Three approaches could be considered in providing FMIS that conforms to smallholder farmers, namely: (1) functionalities they require; (2) the use of free external open data sources; and (3) a mobile apps platform. However, when utilizing external data sources, four main problems may arise, namely: (1) schema heterogeneity, (2) schema granularity, (3) mismatched entity naming or data unit, and (4) inconsistency of data. The ultimate goal of this study is to develop a conceptual model of FMIS for smallholder farmers. This study’s specific objectives are to (1) identify farmers’ information needs, (2) construct an ontology for smallholder farmers, (3) construct a conceptual model (4) create an algorithm to address the data interoperability problems, (5) develop a prototype application, and (6) conduct application testing and impact analysis. The research uses two districts of West Java Province as the case study, with a total of 50 smallholder chili farmers as respondents. Farmers’ Information Needs Assessment (FINA) qualitative data analysis results in the ten most important farmers’ information needs. In addition, according to the in-depth discussion, farmers need five pieces of information. This study proposes an ontology called OntoFMIS specifically designed for smallholder farm management information systems. The general conceptual model of FMIS consists of five layers, namely: (1) farmers’ information needs, (2) assessing the quality of external data sources, (3) extraction of the external data sources, (4) split-match-merge, and (5) presentation layer. The split-match method layer includes two groups of algorithms: one algorithm to extract data from external data sources and another algorithm to address issues with data interoperability. The algorithm to extract external data sources includes four different types of data extractors. In addition, this study employs the combination of four matchers’ methods and five similarity-matchers’ algorithms to transform and load data into the application database. Additionally, this study uses a split-match-merge method to speed up the matching process. The prototype application, called SIMUSTI, is downloadable from the Google Play Store. The result of black-box testing shows that 90% of testers could execute all scenarios without any problems. Furthermore, the user-experience testing showed that the SIMUsTi Android application received a positive evaluation for all categories. The usability testing shows that the respondents do not face any difficulties in using the application. According to the impact analysis, the respondents used SIMUsTi apps besides traditional channels to find information. Regarding the financial aspect, the analysis shows a slight increase in average total income for a single crop production cycle, although it is not statistically significant. The proposed conceptual model is the first to explicitly apply to smallholder farmers based on a data interoperability framework, which is the contribution of the research. In terms of the practical contribution, by using SIMUsTi, farmers were able to find the information they did not previously have, namely other locations growing the same crop

In recent years, teaching-learning methods have emerged into a completely new dimension from what used to be a traditional approach. The in-person lectures have been converted into online virtual learning, the traditional record-keeping has been replaced by robust learning management systems which have made the teaching learning process lot more efficient and convenient. However, the increased use of digital storage and access to students’ personal information raises the concern about data privacy and security for both LET (Learning, Education and Training) users and providers. The lack of knowledge of the users about their own rights and privacy leading to the lack of practice has resulted in illegal processing and piracy of personal data. Thus, this study focuses on modelling the ISO 29187 standard pertaining to LET users and providers and proposes an ontology framework for the standard. Furthermore, the study elaborates on hosting the proposed ontology to establish APIs to create a compliance question-answering system related to LET using a faceted search for the stakeholders of LET.

Sleep apnea is a common sleep disorder that can have significant negative consequences, including cognitive disabilities, excessive sleepiness, and depression. However, it is often not diagnosed, which causes delays in treatment. Standard diagnostic procedures, polysomnography, are complex and expensive and require specialized facilities and personnel. To address this problem, various wearable devices have been proposed to diagnose sleep apnea at patients’ homes. This study aimed to identify the most suitable model for sleep apnea detection using wearable devices. Pure Health LIFE HR 2 and Wellue O2 ring were the wearable devices used in this study. Additionally, polysomnography was used as the ground truth. In the study, patients were required to wear both devices, including polysomnography sensors, while sleeping at the Thammasat Hospital sleep lab. A custom application was used to collect SpO2 (oxygen saturation) data from the smartwatch. The study found that the Pure Health LIFE HR 2 was unable to read SpO2 below 95, while the O2 ring had good accuracy, with a deviation of 5% from the gold standard for SpO2 reading. The data collected from the O2 ring were then used for prediction purposes with a rule-based approach model. However, only this model, which considered an event in which SpO2 dropped below 3% for 10 seconds, achieved only 30% accuracy in predicting apnea. To improve precision, SVM (Support Vector Machine) and 1D CNN LSTM (Convolutionary Neural Network with Long-Term Memory) models were implemented, incorporating additional features such as motion, pulse rate, and patient demographic. The accuracy of the model improved implementing this parameter since only SpO2 data can be misleading to the model. Some apneic patient doesn’t have the oxygen desaturation. The 1D CNN LSTM model had overfitting issue due to the limited and unbalanced set of data containing only 129 patients. The overall accuracy of the 1D CNN LSTM model was 62%, but it tended to predict most cases as positive. Based on these challenges, SVM was chosen because it works well with smaller datasets like ours. The SVM model was found to be 88.8% accurate in detecting sleep apnea, a significant improvement compared to previous models. These results highlight the potential of the Wellue O2 ring as a promising device for the detection of sleep apnea, especially in areas where polysomnography facilities are not widely available.

Sleep apnea, a common sleep disorder, can cause cognitive impairment, excessive sleepiness, and depression. Despite its prevalence, it is often under-diagnosed, leading to treatment delays. Polysomnography, the standard diagnostic procedure, is complex and expensive, necessitating specialized facilities and personnel. To address this, wearable devices have been proposed for home-based sleep apnea diagnosis. This study investigated the accuracy of a smartwatch in diagnosing sleep apnea compared to polysomnography. Fifteen patients wore the devices while sleeping at Thammasat hospital sleep lab. A custom application collected SpO2 (oxygen saturation) data from the smartwatch. The study found that the smartwatch had limitations in reading SpO2 below 95, while other wearable devices showed better accuracy with a 5% deviation from the gold standard. The study recommends further research using machine learning and deep learning for sleep apnea detection using wearable devices. The development of mobile applications and web platforms to extract and analyze vital sleep health data from smartwatches, such as sleep time quality and snoring levels, could provide accessible, cost-effective tools for preliminary sleep apnea diagnostics. These innovations aim to offer user friendly solutions for patients and healthcare professionals to monitor sleep health and make informed decisions.

Nowadays, the changing nature of work in the labor market due to technological advancements has put pressure on workers and organizations. Upskilling and reskilling have become common to ensure workers possess the necessary skills. Graduates often find themselves working in fields unrelated to their education, but the availability of digital materials and online education has made it easier to acquire the required qualifications. However, the vast number of online courses makes selecting the right curriculum challenging. Recommender systems have been proposed to assist with course selection but mostly focus on formal education. This study suggests further research on recommender systems for non formal education. The Thailand Professional Qualification Institute (TPQI) has developed the Professional Qualification Framework to assess and certify individuals’ competencies. Integrating this framework with other systems, such as ontology, can facilitate knowledge representation and connection between systems. However, in Thailand, there is a lack of machine-readable representation for the professional qualification framework and occupational standards. Therefore, this study aims to develop a course recommendation system that utilizes the professional qualification framework and occupational standards, using an ontological framework to match occupations with related courses.

Agriculture is undoubtedly the largest provider of livelihoods in most of the developing country. Also, for most of these countries, the agriculture sector is a significant contributor factor to the country’s Gross Domestic Product. In the context of Nepal, agriculture is the main occupation of Nepalese people. The population is increasing, and so is the need for food. However, the productivity in Nepal is not enough for the Nepalese people. Farmers in developing countries like Nepal face various environmental and economic constraints, such as limited arable land, global warming, infertile soil, and lack of proper resources. The overall motive of the research is to apply the best Machine Learning model to suggest suitable crops to farmers based on their location’s environmental and economic conditions, which results in improved crop productivity. The study uses regression models to predict crop price and crop yield while employing classification models to select appropriate crops based on soil and weather conditions. The research objectives include designing and developing machine learning models for crop selection, comparing different models to find the most accurate crop selector model, developing yield and crop price prediction models, integrating crop selection with production and crop price predictions, and testing the system for functionality. The models utilize various features, including soil and weather parameters, to predict crop prices and yields, and suggest suitable crops. The study employs multiple regressor and classifier models, including Linear Regression, Support Vector Regressor, DecisionTree Regressor, RandomForest Regressor, and XGB Regressor for prediction while it uses KNN, DecisionTree Classifier, RandomForest Classifier, and XGB Classifier for crop classification, with their default values initially. For Crop price prediction and Crop yield prediction, SVR and XGBooster performed well with a test accuracy of 91.6% and 96.4% respectively. For crop selection, Randomforest Classifier performed better with a test accuracy of 94.13%. The models are implemented in web applications created in the Django framework. The results of this study can assist farmers in Western Nepal to make informed decisions about crop selection and production, based on both environmental and economic factors. Keywords: Crop Recommendation, Crop Price Prediction, Crop Yield Prediction, Machine Learning, KNN, Linear Regression, Decision Tree, SVR, XGBooster, Random Forest.

Handwriting recognition technology plays a pivotal role in the digitalization of administrative records, a task of paramount importance in the modernization of governmental processes. This thesis delves into the development and implementation of handwriting recognition systems tailored specifically for the Vietnamese script. The unique characteristics and complexities of the Vietnamese script, with its diacritics and complex ligatures, pose significant challenges for automated recognition systems. The primary objective of this research is to design and evaluate a robust and accurate handwriting recognition system for Vietnamese administrative records. To achieve this, we conducted an in-depth analysis of the Vietnamese script’s linguistic features, cultural context, and historical evolution. This analysis informed the development of a tailored recognition algorithm that effectively addresses the script’s intricacies. Through a comprehensive evaluation process, utilizing a diverse dataset of handwritten administrative documents, we assessed the system’s performance in terms of accuracy, efficiency, and scalability. My results demonstrate the effectiveness of the proposed recognition system in accurately transcribing handwritten Vietnamese text into digital format, thereby expediting the digitalization of administrative records. This research contributes to the broader field of handwriting recognition and holds significant implications for the efficient management of administrative records in Vietnamese governmental institutions. The successful implementation of such a system promises to streamline administrative processes, improve accessibility to historical records, and enhance overall efficiency in the public sector. Furthermore, I will conduct a small-scale experiment to compare and evaluate the performance of Apple’s new integrated graphics processing unit (Apple Silicon M series) in executing machine learning training tasks, analysis, and image recognition. This aims to assess whether they are competitive enough with dedicated graphics cards. If they indeed prove to be effective, it may be suggested to apply them in hardware-equipped tasks to optimize the investment costs and management for VNPT’s AI product development. This is because integrated chips tend to require less space and energy, offering potential advantages in these aspects.

Navigating the transformative intersection of artistry and technology, this study ad dresses a pivotal gap in e-commerce, focusing on the enhancement of visual presentation in product listings. Motivated by the influential role of first impressions in consumer be havior, the research is inspired to redefine the e-commerce narrative by integrating cus tom visual styles through innovative AI technologies. The project is driven by a vision to unlock the potential of artistic expression, aiming to create personalized and captivating shopping experiences that resonate with individuality and creativity. Identifying the unmet need for tailored AI solutions in the diverse landscape of e-commerce, the study is inspired to develop, innovate, and evaluate novel approaches for image syn thesis. The motivation stems from the challenges faced by sellers in balancing standard ization with personalization, and the aspiration to contribute both to the practical realm of online retail and to academic research. The objectives encapsulate the development of a groundbreaking approach, innovation upon existing models, and rigorous evaluation through empirical experimentation in Human-Computer Interaction (HCI). The study aspires to empower sellers with creative tools, envisioning a future where imagination and innovation are at the forefront of e commerce. In essence, this research is a harmonious blend of motivation and inspiration, aiming to contribute significantly to the renaissance of artistic exploration in e-commerce and to provide valuable insights and advancements in the application of AI technologies in the digital retail landscape.

This project is aimed at developing a data governance maturity model and assessment for Van Lang University. Van Lang University (VLU) is a higher education institution based in Ho Chi Minh City, Vietnam. It has a student body of 45,000 and around 2,000 in staff number. VLU has invested considerably in digitizing its information assets in the past decade. However, it still encounters significant challenges in effectively using its data to reach operational and strategic goals. This project attempts to research and review current DGMM offerings on the market and develop a model for use at Van Lang University. The resulting model is a combination of the structure of the Stanford Data Governance Maturity Model and the knowledge from the Data Management Body of Knowledge. It features eleven Data Capabilities assessed across four criteria: People, Processes, Tools, and Oversight. The current model was used to conduct the initial data governance maturity level at VLU by the model developing team and is planned for use institutionally later in the year.

Currently, there are too many television programs airing, making it difficult for viewers to choose. With hundreds of different programs, viewers have countless options, but this also makes it challenging for them to find their favorite shows. Offering content suggestion serves as a way to tackle the mentioned issue. MyTV’s system, until April 2020, didn’t personalize recommendations. Its first attempt in 2021 faced data and algorithm issues, leading to customer dissatisfaction and risk of revenue loss. This project brings solutions to help MyTV address this issue. Hence, the primary aim of this project is to improve the system’s recommendation algorithm. The aim is to create a more personalized, user-centric experience, enhancing customer satisfaction and ultimately, the potential for increased revenue for MyTV. This will be achieved through detailed data analysis, improved programming strategies, and advanced machine learning techniques. Using the customer data and their usage history supplied to the big data system, the project applies superior data processing methods and machine learning models to present the most appropriate content recommendations tailored to the preferences, habits, and behaviors of every single customer or customer group. This helps customers find the content they need accurately. The recommended content spans across various services offered by MyTV, including television services, film and series services, and comprehensive entertainment services. The content recommendations are also based on multiple criteria, such as behavioral criteria, content criteria, and preference criteria.

Accurate traffic forecasting is a paramount requirement for mobile network operators (MNOs) to efficiently manage network resources, ensure optimal service quality, and plan for future capacity expansions. Nowaday, with the trend of mobile subscribers shifting to using data service more than traditional voice & SMS, the need to accurately predict the amount of data traffic of total network in hourly and daily time is essential. So this research begins by outlining the critical importance of data traffic forecasting in the mobile telecommunications industry and the shortcomings of traditional forecasting methods. It then explores the capabilities of deep learning, including recurrent neural networks (RNNs), long short-term memory networks (LSTMs), Gated Recurrent Units (GRU) and convolutional neural networks (CNNs), in capturing intricate temporal patterns and dependencies within the traffic data. The primary metric used to evaluate the results is the MSE, which is used to calculate the model’s accuracy. We conducted our experiments using data from the VNPT data traffic in 2G/3G/4G domain.

Organizations with a wide range of operations must manage a multitude of services and processes. Every organization must be able to track both the services it provides and its overall performance. Unpredictable issues can disrupt business operations; therefore, preventing and mitigating these issues is critical. The goal of this study is to apply data engineering to one of the 27 modules of Regional Container Lines, Intermodal, and Job System. Data engineering includes, but is not limited to, data profiling, data cleaning, data warehouse design, and incremental ETL. Once this is in place, the organization will be able to analyze data to aid in decision making. ChatGPT, on the other hand, took the world by storm when it was released in November 2022. Several use cases for how ChatGPT can be used in various domains to improve their work have been identified, and several companies have integrated this into their products. As a result, the organization will create its own ChatGPT-based application, RCL-GPT, to support employee training and process standardization, with ChatGPT technology at its core.

This 6-month internship report highlights the journey of enhancing the architecture of the HOTS application. Commencing in February and concluding in August, the intern ship focused on a primary objective, addressing real-time data handling and optimiz ing data management for larger datasets. To achieve these goals, a streaming service with Apache Kafka was carefully designed and integrated into the HOTS application. Rigorous testing, important feature implementation like the schema registry and conflu ent control center and SSL encryption were integrated, and deployment on a dedicated server resulted in a robust Kafka architecture. Code optimization techniques, includ ing batch processing and improved memory management, were employed to overcome bottlenecks in HOTS. The outcomes showed promising results with significant improve ments in running time, RAM usage, and dataframe sizes. Participation in Smile’s LLM Hackathon provided valuable insights into real-world LLM applications. The internship fostered valuable experiences in Python development, streaming platform design, and the importance of code optimization. Overall, this report showcases the successful im plementation of architectural improvements for the HOTS application, contributing to a deeper understanding of software engineering and the importance of optimizing code for optimal performance.

The development of mobility technology makes it possible to collect real-time data about tourists in a given location, including their geographical and temporal behavior. Movement data has become an essential alternative data source used in tourism studies. Taxis are an im portant mode of transportation employed by tourists visiting a new city. However, the main challenges of using taxi GPS data in the tourism domain are a lack of semantic information regarding trip purpose and user profile that could be used in an in-depth analysis of tourist behavior. This dissertation proposes TOURISTA data platform to manage and process heterogeneous data, including taxi data, social media data, and place data for tourist behavior analysis. We propose a data pipeline that can be scaled in order to process a significant amount of data regarding taxi trajectory and social media, with two objectives. The first objective is to ex tract the tourist trajectory data from the raw GPS data. This study proposes the TOURISTA model based on a rule-based and probabilistic model to infer the purpose of tourist trajecto ries based on activity and expenditure, considering origin-destination locations. We enhance an existing probabilistic model by building the model using a variety of data sources, includ ing taxi trajectory data, social media data, and place data. The second objective is to extract tourist activities/points of interests (POIs) from geo-tagged Twitter data. We examine ac tual tourist activity from social media data to build an activity popularity model, integrate it with trip and place information, and infer tourist trips using the probabilistic model. For the experiment, we investigated the TOURISTA model by comparing the activity proportions of three baseline methods with results of the Tourism Authority of Thailand’s (TAT) tourist behavior survey for five activities: FoodAndDrink, Spa, Nightlife, Religious/cultural, and Leisure. The results of the proposed method closely match the survey data for several activity categories. We applied data analysis techniques to a case study during the Songkran Festival in Bangkok to reveal tourists’ travel characteristics and activities, tourist movement behavior, and pop ular tourist destinations. The analysis results show that our study is useful and helps under stand tourist flow and the high density of tourist locations at different times, which is crucial for planning or marketing tourism campaigns targeted at specific tourist groups. Govern ment officials and tourism businesses can use this information to better plan or market their tourism campaigns.

The high non-progress rates of students in introductory programming courses have been continuously reported and become a persistent issue worldwide. Educators have been striving to improve instructional design and delivery methods aiming to en hance students’ comprehension and ultimately increase their success rate. Neverthe less, it is essential to recognize that each class has its distinct characteristics, and while good instructional design and delivery are crucial, they can only contribute to part of the overall success. Instructors require competent intervention strategies to handle and elevate unforeseen situations that may arise in each course. This, in turn, presents another challenge of implementing effective intervention. Most instructors struggle to identify at-risk students to determine a proper intervention approach, to trace and to evaluate whether the intervention works. This process requires extensive effort and commitment, which is impractical, especially for large classes with few instructors. This dissertation proposes a platform, namely i-Ntervene, that integrates a Learning Management System (LMS), an automatic code grader, and learning analytic fea tures which can empower systematic learning intervention for large programming classes. The platform iteratively assesses student engagement levels and subject un derstanding to identify at-risk students and suggests aspects of intervention based on their lagging in these areas. Students’ post-intervention data are traced and evaluated quantitatively to determine effective intervention approaches aligning with the evi dence-based research design. The i-Ntervene prototype was experimented with a Java programming course, delivered to 253 first-year undergraduate students. The result was satisfactory, as the instructors could successfully perform and evaluate 12 interventions throughout the semester with minimal administrative efforts. In addition to the proposed platform, the research introduces a longitudinal analysis to investigate the underlying issue of the high student failure rate in introductory programing courses. It examines the student-related factors that differentiate stu dents who failed the course for the first time and those who repeatedly failed from their successful peers, based on the longitudinal observation. These factors encom pass various aspects including student’s demographics, learning motivation & strategies, and engagement levels in learning activities. In this study, we analyzed three consecutive semesters of student data. The longitudinal analysis highlights three common learning motivations and strategies that should be enhanced for all non-progressing students: (1) self-efficacy, (2) effective time & environment man agement, and (3) the perception of the course’s importance and utility. For new stu dents enrolling for the first time, the results indicate that instructors should give pri ority to strengthening their intrinsic goal orientation and control of learning belief, while repeating students should concentrate on promoting their metacognitive self regulation. In terms of learning activities, instructors should primarily emphasize activities that involve coding practices, such as assignments and in-class exercises. Additionally, assistance should be easily accessible for students who encounter chal lenges during coding practices. The findings from the study provide a valuable con tribution to the research community by precisely addressing the issues of high student failure in introductory programming courses.

The high non-progress rates of students in introductory programming courses have been continuously reported and become a persistent issue worldwide. Educators have been striving to improve instructional design and delivery methods aiming to en hance students’ comprehension and ultimately increase their success rate. Neverthe less, it is essential to recognize that each class has its distinct characteristics, and while good instructional design and delivery are crucial, they can only contribute to part of the overall success. Instructors require competent intervention strategies to handle and elevate unforeseen situations that may arise in each course. This, in turn, presents another challenge of implementing effective intervention. Most instructors struggle to identify at-risk students to determine a proper intervention approach, to trace and to evaluate whether the intervention works. This process requires extensive effort and commitment, which is impractical, especially for large classes with few instructors. This dissertation proposes a platform, namely i-Ntervene, that integrates a Learning Management System (LMS), an automatic code grader, and learning analytic fea tures which can empower systematic learning intervention for large programming classes. The platform iteratively assesses student engagement levels and subject un derstanding to identify at-risk students and suggests aspects of intervention based on their lagging in these areas. Students’ post-intervention data are traced and evaluated quantitatively to determine effective intervention approaches aligning with the evi dence-based research design. The i-Ntervene prototype was experimented with a Java programming course, delivered to 253 first-year undergraduate students. The result was satisfactory, as the instructors could successfully perform and evaluate 12 interventions throughout the semester with minimal administrative efforts. In addition to the proposed platform, the research introduces a longitudinal analysis to investigate the underlying issue of the high student failure rate in introductory programing courses. It examines the student-related factors that differentiate stu dents who failed the course for the first time and those who repeatedly failed from their successful peers, based on the longitudinal observation. These factors encom pass various aspects including student’s demographics, learning motivation & strategies, and engagement levels in learning activities. In this study, we analyzed three consecutive semesters of student data. The longitudinal analysis highlights three common learning motivations and strategies that should be enhanced for all non-progressing students: (1) self-efficacy, (2) effective time & environment man agement, and (3) the perception of the course’s importance and utility. For new stu dents enrolling for the first time, the results indicate that instructors should give pri ority to strengthening their intrinsic goal orientation and control of learning belief, while repeating students should concentrate on promoting their metacognitive self regulation. In terms of learning activities, instructors should primarily emphasize activities that involve coding practices, such as assignments and in-class exercises. Additionally, assistance should be easily accessible for students who encounter chal lenges during coding practices. The findings from the study provide a valuable con tribution to the research community by precisely addressing the issues of high student failure in introductory programming courses.

This study investigates the effectiveness of the YOLOv8 object detection model in enhancing hotspot anomaly detection within solar farms, with a primary focus on improving maintenance processes’ efficiency and accuracy. The examination encompasses a comprehensive analysis of challenges associated with hotspot detection, delves into the architectural intricacies of the YOLOv8 model, and outlines tailored training procedures. The resulting optimized model, YOLOv8-OPT, demonstrates significant improvements, boasting a 3.3% increase in precision, a substantial 12.7% improvement in recall, and commendable progress in mean average precision (mAP50 / mAP50-95) by 6.8% and 8%, respectively. However, in comparison to a preceding research model operating on a more constrained dataset, YOLOv8-OPT reveals certain limitations, reflecting a 61.3% decrement in precision and a 53.2% reduction in recall. In this comparative landscape, YOLOv8- OPT emerges as a preeminent model, showcasing superiority across various metrics. This includes a 5.08% increase in precision, a 6.90% increase in recall, a remarkable 16.67% increase in mAP50, a notable 38.89% enhancement in mAP50-95, and a 5.08% increase in F1 Score. The accompanying ablation study underscores the pivotal role of hyperparameters, emphasizing the need for meticulous tuning to achieve optimal model efficacy. While YOLOv8-OPT signifies advancements, it underscores the importance of carefully considering dataset nuances and the imperative for ongoing optimization in propelling future developments in object detection methodologies.

Academic certificates have long been issued as physical paper certificates, which are subject to various problems such as document forgery, loss, and theft, including a slow and challenging verification process. To cope with the mentioned problems, Thai Higher Education Institutes (HEIs) recently started to focus on digital signatures for digital degree certificates. Nevertheless, the chosen technology still often requires contacting the issuing organization to confirm the origin and authenticity of documents which is time-consuming and often tedious. A new standard for digital credentialing called Verifiable Credentials (VCs) has recently been ratified by the World Wide Web Consortium (W3C). The approach enables a more portable, privacy-preserving, and user-centric digital identity that can be adopted on a large scale through cryptographic and Blockchain technology. This research work aims to adopt the VCs standard for implementing the Verifiable Credentialing ecosystem for the HEIs. The ecosystem composes of web UI for both the issuer and verifier to create, issue, remove and verify digital credentials through the controller component of the Hyperledger Aries agent. Built-in business logic is used to manage HEIs credentialing process. An issuer uses the web UI to issue and offer certificates to the holder’s e-wallet. Blockchain is used to store information needed to validate the credential’s authenticity. Verifier can request required credentials in the form of presentation proof from the holder. With the built-in Zero-Knowledge Proof (ZKP) function, the holder can choose to share full or partial information that is relevant to the verifier.

In addition to logical reasoning, computer programming is seen as a crucial talent for the development of problem-solving abilities. A number of basic programming environments have been developed to assist newbies in avoiding typical programming errors such as gram mar and logic. Most of these breakthrough environments use block-based programming languages rather than inputting instructions for example, Scratch which the instructive and entertaining activities offered and a free and open-source programming language. However, computer programming is a difficult course to teach and learn. the learning process differs for different learners, depending on their personality, cognitive abilities, and past learning experiences. To support student’s differences, an personalized and adaptive mechanism are need to generate individual suggestions. In this thesis, students were assigned practice exercises and suggesting related learning resources based on the characteristics and performance. More over, the learner profiles, shareable and composable scratch learning resources and learning path were modeled from comparing various metadata standard models to support above mechanism. For recommendation, Rules-Based Personalization were used for mapping the student personal profile and learning object. The evaluation of this thesis are divided into two parts: student experiment and teacher feedback from interview. The purposes experiment are measuring the 1) Educational met rics or Knowledge assessment which comparing average performance improvement between two groups; 2) User-centric by filling USE questionnaire (satisfaction) and 3) calculating the accuracy of suggestions of ScractchXChange by using teacher’s opinion. As a result, ScractchXChange can recommend learning resources that meet with student need as students satisfied with the system as well as all students have gain score. In addition, the average accuracy suggestions Personalization Module from teachers is in good level accuracy.

Conversational bots in education have been growing extremely popular day by day. These programs are utilized as alternate to direct teacher for lesson and also able to track and store student actions throughout the learning process. Researchers are growing intereste on this data for learning analytics and creating student behavioral prediction models or student profiling. Thailand has used Scratch in basic education to boost computational thinking ability of the kids. But UNESCO has detected shortage of suitable instructor in many provincial institutions in compared to number of students, for which ScratchThAI chatbot was launched to aid students directly by sending learning materials, exercises, as well as other helps. The major purpose of my study is to allow instructors to enhance their intervention plan and teaching effectiveness by recognizing distinct group of students. During a session designed to introduce ScratchThAI to three distinct educational institutions, two types of data were collected for this study: chatbot interaction and performance data. Then, exhaustive experiments were undertaken to establish the most effective clustering approach in terms of performance evaluation and clustering quality. Five student groups were found using the K-means clustering technique, which yielded the best results in terms of classifying students according to their characteristics. Later, a model for early prediction was developed employing clusters as the model’s classifier to identify various student groups in advance. Based on several performance matrices, the XGBoost classification model yielded the best outcome. For teachers to readily comprehend the features of different student groups and individuals, a visualization dashboard was constructed. Finally, I suggested implementing the selected prediction model into the dashboard so that teachers can identify student groups earlier in the semester.

Visiting museums provides various benefits to visitors, such as providing intriguing history, culture of a certain location, and inspiring visitors via displays. In recent years many public sites have been forced to close temporarily due to the coronavirus epidemic. Furthermore, there are other barriers that may hinder tourists from physically visiting the museum, such as health issues, geography, and time constraints. So, web-based technologies can assist in improving the design of online learning environments for museum. Due to large amounts of data might make it difficult for a user to locate interesting stuff. As a result, the goal of this study is to create an online tour platform for Museum Pool, where users may browse by type of museum and the platform would suggest related artefacts based on recommendation systems. Furthermore, the result accuracy of classification type of museum still moderated. And the recommended still need to evaluate by user satisfy rating on website.

Business intelligence (BI) improves data-driven decision-making. BI systems aggregate data from various sources and display a matrix. BI technologies need self service data visualization to let users’ access and understand data fast, even for flexible questions. ISPs must manage data silos and retain massive amounts of network data and usage data. When distinct departments are responsible for incomplete datasets, other departments don’t know what data exists in their firm. Data silos often cause chaos of different data definitions and inconsistency. These obstacles slow cross functional BI data integration for enterprise-wide analytics. Most reviewed papers focus on system architecture or large data analytics, there is a gap for research contribution on the empirical study of BI and analytical system governance and architecture framework implementation on enterprise BI. This work aims to propose an architectural framework for BI in the ISPs, which also could be used as a guideline for the practitioners in other domains to achieve satisfied BI and analytics which provides effectively, efficiency, trustable, and sustainable to the business and IT infrastructure. First, analytical system architecture and data pipeline from previous works are summarized. Second, the potential solutions for coping with data silos are discussed. Third, several techniques for query speed improvement are investigated. Next, a flexible cross-functional architectural framework for BI in the ISP is developed to enable fast and flexible analysis, supporting different groups of users in the enterprise. Lastly, the BI prototype has been implemented on an ISP firm to evaluate the successes of the proposed framework. The result system reflects the importance of data integration automation with multiplatform architecture improves the centralized-BI and analytics system efficiency. Besides, collaboration between IT and businesspeople under enterprise’s data governance helps provide trusted data to users.

The improper tag organization has been derived by data providers who provide data categories and tags for a dataset to be published on the ThOGD portal. They have currently guided by the available autocomplete function in the portal. With this application, data categories and tags to be suggested to data providers are forecasted from the historical data that was provided by previous data providers. This results to a consequence of several datasets with similar contents but are labeled with different tags in similar meaning are found in the portal. Besides, data consumers cannot get the information being matched to their preference according to the filtering of data category and tag. In this study, our contributions for overcoming the above-mentioned challenges have two main sections, including the attention-based categorical identifier and the topic hierarchy-based categorical concept hierarchies. With the use of Attentive Deep Supervision, there is a weighted effect on loss optimization of the categorical identifier. With the use of Topic Hierarchy, Latent Dirichlet Allocation (LDA) topic modeling is utilized for potential tag term extraction, Heterogeneous Evidences are exploited for relation identification, and Anytree is employed for hierarchy construction. By applying these approaches, the macro average of precision and F1-score of the attention-based identifier improves by 0.6640 % and 0.5570 %, respectively. The micro average improves by 0.8060 %, and 0.6980 %, successively. Meanwhile, the concept hierarchy based categorical concept hierarchies can provide comprehensive tags related to a dataset to be published because of the recommendation strategy that assigning tags with the same highest important weight to the same rank.

Computational Thinking (CT) is officially introduced into the National Curriculum worldwide. Scratch, a block-based visual programming language, is widely adopted in primary-level schools. However, traditional learning support systems lack important capabilities to support CT teaching and learning in a regular classroom. Recent CT development through programming in formal education has presented three important challenges: P1: Mixed-ability students in the same class, P2: a high student-teacher ratio, and P3: a learning-hour limitation. To effectively address the challenges, this research proposes ScratchThAI+ framework to enhance CT teaching and learning through Scratch programming. The framework provides a kid-friendly chatbot and precision education features (i.e., assessment, treatment, and prevention) to the Scratch website. The chatbot assists diverse students in learning and practicing online coding anywhere, anytime. The assessment analyzes students’ learning activities and products throughout their practice. CT competency and difficulties are recognized for both individuals and the class as a whole. The treatment and prevention feature then initially intervene by suggesting extra exercises and learning materials, to fill the gaps. The framework provides a dashboard to visualize students’ learning progress and learning performance. This enables teachers to identify at-risk students earlier and manage appropriate interventions within the constraints of limited school hours. The research is organized into three phases with three objectives: (1) understanding traditional classroom practices and challenges in CT development and assessment, (2) designing and developing a teaching and learning support framework for precision CT education, and (3) developing a precision education-based assessment by adopting evidence-based approaches. The first phase explores Thai school challenges, importance, and strategies using case studies and a survey. The case studies reveal teachers’ opinions on challenges and strategies. The survey determines the top five challenges. Then, we address the three challenges (P1-P3). The second phase designs and develops ScratchThAI, a conversation-based learning support framework, which adopts relevant educational approaches and relevant technology-enhanced learning (TELs) in its design and architecture to address the challenges. ScratchThAI is adopted as part of Grade 5’s computing science course in a Thai primary school. The results reveal that it can alleviate the challenges with the following supports: achieving better learning performance, enhancing teaching, and learning activities, promoting student engagement and motivation and positive teachers’ and students’ satisfaction. The third phase designs PLUS assessment model which adopts an evidence-based approach into its design and integrates with ScratchThAI, called ScratchThAI+. ScratchThAI+ collects evidence throughout students’ practice (i.e., evidence of the learning process and final products in problem-solving phase and programming phase) and analyzes it against a predefined rule-based system. This enables the provision of accurate and timely assessment information, which is hard to achieve in traditional classroom practices. We investigate ScratchThAI+ effectiveness in an online workshop context and compare its assessment results with traditional automatic assessment. The acquired results present more accurate grading, precise identified mastery and difficulty, earlier identify at-risk students, and provide better understanding of students’ learning. Thus, ScratchThAI+ contributes as a possible and practical solution to CT development and CT education improvement under the aforementioned challenges and also illustrates benefit of adopting precision education in practice.

The prospect value creation of open government data is activated once the reuse of the data transpired. The controversial data over the wall model is prevalent as reflected by open as default in many countries. Governments publish the data as the default on central portals. The majority of the dataset on open data platforms were evidently presented low usage. The contradiction between provision and consumption is insufficiently comprehended. Realiza tion of relationship and contradiction, as well as the status of the usage can support the policy-making process to advance the open government data project. The analytical framework using the business intelligence approach was developed to analyze the open government data portal usages and user behaviors to understand such differences and relationships to finally recommend the improvement to reach the goal in advancing the open government data project. The three research questions are as follows: (1) “What are the provision and consumption patterns?”, (2) “What are the most popular search terms, failed search terms, and what are the reasons causing failed search?”, and (3) “What are the user requirement and feedback that were unfulfilled by the provision circumstances?”. The data pipeline and interactive dashboard were developed accordingly. The results showed which data groups needed to be improved on numbers of published datasets and update frequency. There are associations between the published datasets, the view frequency, and the update frequency. The high-value (HVD) datasets dominated the most views over the non-HVD datasets. The data files with the most prominent usage are the machine-readable formats (mostly XLSX and CSV). The failed search was caused by both data unavailability, tag unmatched, and portal technical problems. Data unavailability or non-discoverability is the most noted data request motivation (54.84%). Shapefile is the most requested format (37.93%). The prototype system evaluation acquired positive feedback on its usefulness; however, the usability is expected to be ameliorated. Two main contributions bestowing in this study are (1) analytical framework validation re sults for portal usage and user behavior analysis that elucidated an insight into the usage in the actual environment, and the result-based practical recommendation, and (2) prototype business intelligence system for CKAN-based open data portal as the practical contribution.

Applications of chatbots in various sectors have been a trend since around a decade ago. Among diverse sectors, education is one of the sectors that use chatbots for many purposes, ranging from answering questions about the institution to checking students’ homework. Ac cording to previous studies and investigations of academic barriers of young learners in this study, it is still lacking in the area of answering students’ questions related to subjects’ con tent due to unbalanced student-teacher ratios in classes. Students have to wait their turn to get their answers, and teachers do not have enough time to answer all the students’ questions. Moreover, students do not get individual support due to big and overcrowded classes. There fore, the overall objective of this study is to develop a framework that can answer primary students’ questions. The study was conducted in five steps: expert interview, requirement specification, system design, development, and evaluation. The first step was conducting two interview sessions to identify academic barriers of primary students and questions types that are frequently asked by primary students. Based on the experts’ input, academic barri ers, functional and non-functional requirements were defined in the second step. In the third step, the most critical part of this study, Subject Ontology(SO) was modeled, and the system architecture of the SubBot framework was designed as well. SubBot framework was devel oped and implemented in the fifth step, system development. Lastly, eight primary teachers evaluated the SubBot framework. All the suggestions were positive, even though some users believed that some students might struggle using SubBot. They concluded that at least 85 percent of students will feel satisfied with the answers returned by the SubBot bot. Our contributions in this study are twofold. Firstly, we modeled Subject Ontology(SO) for the knowledge base of the SubBot framework. Secondly, we developed the SubBot framework for supporting primary school students. In the SubBot framework, a teacher can manage subject knowledge so that SubBot can answer students’ questions.

Institutional digital assets are the foundation for a digital repository (IDR) to provide services for the researchers. As the publishing of these digital assets is growing rapidly, the efficient way of managing and visualization of them is becoming challenging. Several limitations in user interfaces depend on the discovery of resources, document centered design, and approach of searching limit the widespread use of the previous services that IDR provides. Academics and experts who need to read and handle proper documents require things that support them to identify and manage this flood of information. In this study, we have tried to overcome the above-mentioned challenges by proposing an interactive web-based IDR’s visual plugin aimed at supporting the usability and retrieval of institutional digital assets. With the use of Latent Dirichlet Allocation (LDA) topic modeling, visualization technique is utilized for topic analysis in this study. Topic modeling helps for grouping a set of digital assets into several clusters. We also adopted a method of best topic selection for deciding which terms to present to a user to aid in the task of topic interpretation. Sunburst, Treemap and Word Cloud visualization techniques are adopted to exhibit the scholarly resources. Our plugin able to boost searching in scholarly resources by delivering information about the range and context of digital collections and enabling users to more easily browse and explore the contents. Besides, allows user to upload their own bibliographic data for modeling and visualization. The plugin has been applied to model a case study of CSIM digital repository, which can manifest its feasibility and effectiveness in managing and visualizing institutional digital assets.

Dealing with plants in farms is one of the important challenges during the cultivation. Plants can be damaged by various factors, especially plant diseases and insects. When they are damaged, different kinds of abnormal characteristics can be occurred and observed by farm-ers. In real-life situations, farmers manually manage the observation data by comparing it with their background knowledge; or consulting the others e.g. friends and agronomists. To improve this traditional practice and to avoid the problems of diverse skills and personal knowledge, we have investigated and proposed a novel framework to manage observation data. Specifically, we are interested in managing multiple observation data from one or various farmers at a certain time and place since a plant disease can be disseminated. Currently, various knowledge related to plant cultivation has been widely published on the web. This has motivated us to externalize and utilize such knowledge for detecting of plant diseases based on the observation data and recommending appropriate treatments. To this end, both observation data and existing knowledge are investigated in theoretically and prac-tically. In the theoretical phase, we formalize desirable characteristics of observation data from farmers and develop a farmer’s observation management formalism which deal with an observation data representation, treatment representation and, observation data composi-tion. Key characteristics of the formalism are defined according to a type for representing observation data and a composition function for composing relevant observation data. We also show that our formalism helps any observation data to become richer and improve a mechanism of recommending treatments for plant diseases. In our practical phase, we have applied the proposed formalism into a practical use case of rice plants in Thailand. We study: (1) a representation of observations data, in which it can be deployed in practice at Thailand rice fields, (2) a representation of knowledge base of rice plant, which enables to refer rice diseases and treatments by a farmer’s observation, and (3) the design and development of our expert system reflecting our proposed framework from theoretical to practical studies. Our contributions in this part are twofold. First, we model ontologies using the Web Ontology Languages (OWL) for Rice Disease Ontology (RiceDO) and Treatment Ontology (TreatO) from reliable knowledge sources. Second, we develop RiceMan which is an expert system for supporting daily-life activities of farmers and agronomists who are responsible to advise farmers. In RiceMan, users can opt in to integrate their observations with others for the disease and treatment prediction. This composition mechanism (coincided with our theoretical formalism), together with ontology reasoning, lies at the heart of RiceMan.Finally, our thesis is evaluated practically with four kinds of stakeholders: (1) domain expert agronomists, (2) non-domain expert agronomists, (3) students in agricultural field, and (4) ontology engineers. Results are carefully analyzed and discussed immensely in this thesis.

As Information Communication Technology (ICT) has progressively become the world’s favored means of handling digital information, many museums, libraries, and cultural heritage centers have invested in documentation of their collections and published their materials via the Web, making them available throughout the world. The challenges of maintaining valuable cultural artifacts include archiving, managing, integrating, and distributing cultural heritage 3D models using ICT technology. This well-organized data model supports collaborations, sharing of museum resources, and descriptions of cultural artifact resources, enabling self-development experiences and distribution of museum knowledge via 3D virtual exhibitions. This academic work discusses an integrated system to extract 3D cultural heritage objects from repositories. This system is particularly useful for building 3D virtual exhibitions as well as archiving and viewing the 3D cultural artifacts on a Web browser. It uses WebGL technology, and collaborating museums are organized homogeneously by using Linked Open Data (LOD) technologies. Metadata and ontology standards are applied to support the entire curator content management process, including 3D object digitization to improve structured searches. A distributed platform and a 3D virtual exhibitions system using a Resource Description Framework (RDF) enhance information access, semantic searches, and shares, and link museum resources to virtual exhibitions. An authoring tool developed using Unity3D can organize 3D cultural artifact models, set virtual exhibition layouts, and present virtual exhibitions in a Web browser without installing any additional plugins. This research work was based on the implementation of distributed platforms and 3D virtual exhibition systems in the Chiang Mai National Museum, the Hariphuchai National Museum, and the Chiang Saen National Museum in order to analyze collaborations and investigate the museum systems. Over a six-month period, this experiment analyzed data collected from museum staff and museum visitor surveys before and after deployment of the system. The research found that the proposed online museum system significantly improved the museum staff’s efficiency in organizing museum work. In addition, museum users gained a greater knowledge of history, and post-implementation surveys showed high levels of user satisfaction with the museum systems. As a result of system implementation, education, research, promotion, and recreation for museums and online visitors all benefited.

As the name of the project suggests “Acad-mitter”, is a student-teacher interactive application which helps introducing a university’s academic curriculum by using specific features from various famous social networking sites. Most of the features in this project are inspired by Facebook, twitter, Moodle and the CIS (course information system) that is currently being used by AIT . The aim of this project is to help students connect with the teachers in a much easy and an interactive way. All the details regarding the registration, courses, grades, attendance, lecture slides and important announcements are covered in this project. Moreover, there is also a fully updated self-sufficient website that everyone has access to. Tasks such as login, check grade, check attendance, create a group, register for the courses etc. are designed as a web application. Microsoft.NET is used to access data stored in an SQL-SERVER database and the project is created in Microsoft Visual Studio Web Developer 2010 Express.