Saint Michael's College of Laguna

Cyberbullying has a significant impact in the mental health of individuals which were targeted in several studies to find answers whether the statement made is abusive or not. As a direct way of providing good and credible facts on cyberbullying, a study which aims to incorporate the assessment of risk levels in cyberbullying messages based on their type such as flaming, outing, denigration, and threatening is conducted. A dataset of 1,000 Tagalog messages was collected from online platforms and classified using both machine learning and deep learning algorithms. Additionally, a risk-level assessment framework was designed to determine the severity of cyberbullying messages. Experimental results show that in Model 1, which classifies cyberbullying messages, Naïve Bayes and Support Vector Machine achieved an F1 score of 58%. Meanwhile, in Model 2, which categorizes messages into four cyberbullying types, Naïve Bayes and Random Forest recorded a balanced trade-off between precision and recall, achieving an F1 score of 64%. Therefore, future work should consider applying ensemble techniques, expanding the dataset, and employing robust text categorization methods to further improve overall performance.

Voice cloning technology has significantly improved over the past few years, enabling the generation of synthetic voices in different languages and allowing for voice cloning. However, AI-cloned voices face limitations in dataset availability for Southeast Asian languages, making it difficult to accurately generate speech patterns due to variations in pronunciation and intonation. This study aims to evaluate AI-cloned voices based on their pitch, tone, and intonation, identifying which best resembles the original voice. To achieve this, 1,200 audio signals were collected from YouTube search queries and underwent pre-processing techniques. The results revealed that Rask performed best in pitch, with a Mean F0 score of 203.07 Hz, indicating excellent performance. In tone analysis, Altered achieved a score of -207.16, which falls within the <5 range, signifying excellent speech recognition. For intonation, Rask recorded a score of 197.16, demonstrating superior accuracy in analyzing vibration rates. These findings highlight the variations in pitch, tone, and intonation due to regional accents and voice quality. This study serves as a benchmark for further optimization, aiming to improve the accuracy and adaptability of AI-cloned voices in diverse linguistic contexts.

Skin color naturally varies among individuals due to genetic differences, but facial discoloration can signal cosmetic or medical issues. Although skin tone is primarily genetic, discoloration can occur due to various factors, including prolonged sun exposure, cosmetic issues, and medical conditions that may require attention. This study aims to assess the severity of facial discoloration across different skin types. To achieve this, 6,042 images were collected and augmented to enhance model robustness in analyzing facial discoloration severity. A cosine similarity measure was employed for identification based on the Fitzpatrick scale. Gaussian blur filtering was applied to improve structural retention across skin tones and preserve image quality. Canny edge detection was employed to segment discoloration regions, which were subsequently used to distinguish between images with and without discoloration automatically. Three machine learning models were tested, with results showing that the Random Forest algorithm achieved the highest accuracy (75

Hand gestures are widely used across various applications, and research on hand gesture recognition continues to evolve due to its usability. However, challenges such as occlusion remain significant. Therefore, this study aims to introduce a novel approach using pose estimation to evaluate distraction levels based on rate, duration, and saliency while addressing occlusion-related issues to enhance keypoint detection accuracy for more reliable hand gesture analysis. To achieve this, a custom dataset was built, and data augmentation techniques were applied to enhance the robustness of the “Palms Open” pose analysis. Additionally, a Kalman filter was implemented to track missing keypoints, while a Generative Convolutional Network (GCN) was utilized to infer them, improving overall accuracy. The results demonstrated that the model achieved a mAP@50 of 97.70

Attention is a valuable tool for assessing learning, and this study proposes a novel approach to measuring learners' attention levels by implementing an experimental approach that uses audio signals based on waiting time, response length, and response accuracy. The study collected 30 audio signals during an online session in which participants answered questions on AI topics. A criterion was designed to assess the attentiveness of the learners. The findings revealed that the level of attention depends on the questions, with approximately 54% of them marked as having a "very high attention" level. Moreover, considering their status, the majority of the learners had a moderate level of attention. Notably, students in the lower level reported a very high level of attention compared to students in the higher level. The findings emphasize the need for educators to consider individual differences in attention levels when developing and implementing teaching processes. Additionally, educators should consider the impact of technology on attention levels and suggest the need for further investigation into its effects on academic performance..