Information Technology Journal KMUTNB

Employ Innovative Algorithms to Assign Weights to YouTube Search Results and Extract Education - Related Data

Jincheng Zhang — 2026-07-20

The rapid advancement of data science, artificial intelligence, and machine learning significantly impacts education, with generative AI technologies like ChatGPT increasingly influencing learners. This study analyzes YouTube video view data to extract educational insights through a weighted approach. We found that videos with over 100,000,000 views received a weight of 30, while those with 10,000 to 99,999 views were assigned a weight of 1. In sentiment analysis using Python, we determined an average sentiment weight of 0.722 for education, indicating a generally positive public perception. Our proposed "Weighted Extraction Weight Addition - TF-IDF" algorithm calculated a TF-IDF value of 0.060 for "education," yielding a final value of 754.16 when weighted with 12,513 views. Furthermore, the "Weighted Extraction Maximum Weight - TF-IDF" algorithm identified "child" as having a maximum weight of 30, highlighting public interest in children's educational content. This analysis aims to enhance educational development by providing valuable information derived from YouTube search results.

Deep Learning-Based Object Recognition and Detection for Identifying Champa

Phan Thi Thanh Thuy — 2026-07-20

Object recognition and detection have experienced remarkable advancements with the emergence of deep learning, enabling more precise and efficient identification of complex patterns. This study explores the application of deep learning techniques for recognizing and detecting Champa-related objects using a dataset of 5000 labeled words. Our approach integrates convolutional neural networks (CNNs) and transformer-based architectures to enhance recognition accuracy and localization performance. The dataset consists of diverse Champa-related objects, including inscriptions, sculptures, architectural elements, and cultural artifacts, making it a valuable resource for training deep learning models. By leveraging CNNs for feature extraction and transformer-based models for contextual understanding, we achieve improved object classification and spatial localization. The proposed model is evaluated on multiple benchmark metrics, demonstrating its effectiveness in identifying Champa heritage with high precision and recall. Experimental results indicate that deep learning-based methods significantly outperform traditional computer vision approaches in recognizing and detecting Champa-related objects. The findings highlight the potential of artificial intelligence in supporting cultural heritage preservation and archaeological research. Moreover, our study provides insights into the role of deep learning in digital humanities, offering a framework for future applications in heritage conservation and historical documentation. By utilizing advanced neural network architectures, this research contributes to the growing intersection of AI and cultural studies. The proposed methodology not only facilitates the recognition of Champa artifacts but also paves the way for broader applications in heritage conservation. Future work will focus on refining the model and expanding the dataset to encompass a wider range of Champa-related elements.

The Development and Evaluation of Perception on Online Loan Education through Motion Graphics

Jakkaphan Raksasawat — 2026-07-20

This study aims to 1) develop a motion graphic to educate viewers about online loans, 2) evaluate the content accuracy and infographic quality, and 3) assess users' perceptions and satisfaction with the motion graphic. The tools used in the study include: 1) a motion graphic developed using Adobe After Effects and Adobe Photoshop with narration provided by the Luvvoice platform, 2) a content accuracy and quality evaluation form completed by three experts, and 3) a user perception and satisfaction survey on the motion graphic for online loan education, completed by 60 participants. The findings reveal that: 1) the developed motion graphic achieved the highest level of content accuracy and quality (x̅̅ = 4.87, S.D. = 0.19), reflecting the highest assessment ratings from experts; 2) user perception of the motion graphic reached 81.3%, indicating a high level of engagement and comprehension, and 3) the satisfaction rating for the motion graphic was also at the highest level (x̅̅ = 4.62, S.D. = 0.46), indicating that the material effectively conveys clear and relevant information, enabling recipients to understand the content and information about online loans accurately.

Modal Deep Learning for Tongue Diagnosis

Niparat Boonkun — 2026-07-20

Tongue diagnosis is a key component of traditional medical systems and increasingly relevant in intelligent healthcare. This study proposes a multi-modal approach integrating visual and textual data to classify tongue images into three diagnostic categories: Normal, HF Tip, and Abnormal. We compare the performance of three deep learning architectures: Convolutional Neural Network (CNN) for image-based classification, Long Short-Term Memory (LSTM) for text-only processing, and a hybrid CNN+LSTM model for multi-modal fusion. A 10-fold cross-validation was conducted on a curated dataset of 1,273 tongue images and corresponding medical descriptions. Results demonstrate that the fusion model significantly outperforms unimodal models with an accuracy of 98.11%, indicating the effectiveness of combining image and language features in automated diagnosis systems.

An Unsupervised Machine Learning Model for Identifying Suicidal Posts Using Latent Dirichlet Allocation and Text Feature Extraction

Natthaphong Suthamno — 2026-07-20

This study aims to enhance the performance of unsupervised machine learning models in identifying suicidal behavior from English-language Reddit posts. Using Python, a model was developed and trained on 232,074 posts with Latent Dirichlet Allocation (LDA) for topic modeling. Text feature extraction techniques incorporating unigram to 4-gram models were applied prior to model input. Four experimental setups tested different feature word sizes: the top 50,000, 100,000, 150,000 most frequent words, and all available features. Results showed that using the top 50,000 features yielded the highest accuracy (85.45%) but the lowest recall (95.55%), while using all features achieved lower accuracy (79.54%) but higher recall (97.70%). The findings suggest that reducing feature quantity enhances model accuracy and focus on suicide-related content, but may compromise recall due to loss of contextual information. Overall, the study concludes that LDA-based models can perform effectively without relying on the full vocabulary of large datasets, though a trade-off between accuracy and recall must be considered.

Full Issue

Information Technology Journal KMUTNB Information Technology Journal KMUTNB — 2026-07-20