This community contains research publication by the university staff and conference proceedings held by the university. http://repo.lib.sab.ac.lk:8080/xmlui/handle/123456789/2 2026-06-03T05:12:25Z 2026-06-03T05:12:25Z Nagodavithana, S.D. Rupasingha, R.A.H.M. Kumara, B.T.G.S. http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/5313 2026-06-02T05:09:38Z 2026-01-28T00:00:00Z

TikTok Video Classification Based on Emotions Using Convolutional Neural Networks Nagodavithana, S.D.; Rupasingha, R.A.H.M.; Kumara, B.T.G.S. Emotion recognition is a product of AI and computer vision, which enables systems to decipher human emotions during a range of uses. As TikTok grows, expressions made by users in the short videos offer valuable information that can be used in real life to study emotions. Detection of emotions in these videos is, however, a challenge because of changes in lighting, angles, and user behavior. Emotion-based TikTok video classification can be essential to enhance content recommendations, interaction, and emotion-driven social media analytics. To overcome this challenge, as an objective of this research, a Convolutional Neural Network (CNN)-based approach is prepared to categorize TikTok videos based on their emotions. This study is novel in its comparative evaluation of CNN-based emotion recognition models on real TikTok videos, an area largely ignored by existing methods. A sample of 4,000 TikTok videos were gathered, containing the main four emotions, namely happiness, sadness, anger, and surprise. Videos were processed during preprocessing, which involved compression, frame conversion with Python and OpenCV, and normalization and data augmentation. The features were extracted, and models were trained and tested using the three CNN architectures, namely, MobileNet, VGGNet, and ResNet. The data was divided using a percentage split, as well as 5-fold and 10-fold crossvalidation, with different epochs and batch sizes to achieve the best model performance. Accuracy, precision, recall, F-score, and error rate were used as a measure of model performance. ResNet was the most accurate with 96% accuracy, which is far better than VGGNet (71%) and MobileNet (58%). These results prove the usefulness of CNN architectures to classify emotions in TikTok videos and how they can be applied in social media analytics and affective computing. Future research will focus on multimodal audio-text methods for real-time emotion analysis on social media.

2026-01-28T00:00:00Z Vitharana, K.S.N. Kumara, P.G.P. http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/5312 2026-06-02T05:01:13Z 2026-01-28T00:00:00Z

A Study of Machine Learning Models for Text-Based Mental Health Prediction in Sri Lanka Vitharana, K.S.N.; Kumara, P.G.P. The widespread use of social media has revolutionized the way people share personal problems, which is a new line of identifying the occurrence of mental illness at an early stage. This approach becomes especially vital in such circumstances as the Sri Lankan context, when cultural stigma is a real obstacle in the process of seeking help. In order to fill the gap of machine learning applications in this field, this paper explores the automated detection of mental health conditions on Facebook posts in Sinhalese. A mental health expert annotated a corpus of 3,096 posts with a multi-label classification schema (Anxiety, Depression, Suicidal Ideation, Irrelevant) to indicate possible comorbidities. A traditional Random Forest classifier and the new transformer-based models, BERT and RoBERTa with explicit hyperparameter settings, were tested and compared to perform this multilabel classification task. The performance analysis indicated that there are serious gaps. The Random Forest model obtained a low efficacy, indicated by its macro F1-score of 0.45, which is poor at predicting the important suicidal ideation class (F1-score: 0.33). This baseline was significantly low in comparison to the transformer models. The BERT model had a strong macro F1-score of 0.83, and the RoBERTa model had the best overall score of 0.85. These findings indicate the superiority of transformer-based models, namely RoBERTa, in this sensitive classification task. The analysis shows that natural language processing has the potential to be successfully used to detect the indicators of mental distress in the specific sociolinguistic environment of Sri Lanka. While limitations include depending on a single annotator and platform-specific data, ethical issues associated with the application in the real world. This study serves as a foundation for developing proactive digital solutions that can enforce mental health surveillance and early intervention, potentially overcoming the stigmatization barrier.

2026-01-28T00:00:00Z Senanayake, S.M.N.D. Rupasingha R.A.H.M., R.A.H.M. http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/5311 2026-06-02T04:26:17Z 2026-01-28T00:00:00Z

Classification of Dental Lesions Using Camera-Captured Images with Convolutional Neural Networks Senanayake, S.M.N.D.; Rupasingha R.A.H.M., R.A.H.M. Oral cavity is a central part of the appearance of a person and their health condition and oral care is crucial. The consideration of good habits and the early identification of lesions are paramount, and their diagnosis is normally performed with the help of visual examination, radiographies and biopsies. While these methods are widely used, they present several challenges. Early lesions are often small or resemble healthy tissue, making detection difficult. Diagnosis is further limited by the subjectivity of visual assessment and the time-consuming nature of radiograph interpretation. Because of these challenges, researchers are increasingly looking at how artificial intelligence can help detect periapical lesions. Many studies have focused on identifying teeth or dental diseases using X-ray images, but research using RGB (color) images is rare. RGB images are easy to capture, non-invasive, and more accessible during routine dental check-ups, making them useful for practical AI-based diagnostic tools. To fill this gap, the objective was to use deep learning model to automatically detect dental lesions and improve diagnostic accuracy. In this approach, we evaluate and compare different convolutional neural network (CNN) architectures for identifying three major dental lesions namely, Gingivitis, Calculus, and Hypodontia from 4000 optical color images captured in front of the mouth. After pre-processing and extracting features, the dataset was trained with three pre-trained architectures: EfficientNetB0, DenseNet121, and ResNet50. The findings indicate obvious variations in performance, and DenseNet121 has always got the maximum accuracy of 86.91% and higher precision, recall values, f-measure values compared to other models. The future dental industry may benefit from this research as it will be easier to detect issues early and offer cheap equipment to improve oral health. The research compares the CNN performance on dental lesion classification and prepares the way to predict the severity and medical application in the future.

2026-01-28T00:00:00Z Pathiranage, W.P.T.N. Kumara, B.T.G.S. http://repo.lib.sab.ac.lk:8080/xmlui/handle/susl/5310 2026-06-01T09:56:20Z 2026-01-28T00:00:00Z

Deep Learning Framework for Predicting LinkedIn Follower Count Range Pathiranage, W.P.T.N.; Kumara, B.T.G.S. LinkedIn has become a key platform for professional networking, where follower count increasingly reflects visibility, credibility, and digital influence. Existing research offers only limited insight into how different factors jointly shape follower growth. Insights from other social media platforms do not translate well to LinkedIn’s professional context. This study examines factors influencing LinkedIn follower count using a multiinput deep learning model. The model integrates three major data modalities professional, demographic, and facial-emotional features allowing a comprehensive multimodal prediction approach. The study focuses on three key areas: determining whether a 1D CNN based multimodal model performs better than classical machine learning models; identifying which feature groups most strongly influence follower count; and evaluating the extent to which a multi-input 1D CNN can learn complex non-linear interactions more effectively than traditional approaches. Structured and LinkedIn profile data and facial-emotional indicators obtained based on profile images were used in classification. Standardized cleaning, one-hot encoding and MinMax scaling were used to process features. The classical models used were compared to a 1D CNN with a multi-input. Accuracy, F1- score, MAE, MSE and Explainable AI techniques were used as model evaluation. According to the results, LinkedIn user following is mainly motivated by career advancement and profile display and not by demographical factors. The research conclude that multimodal deep learning is a highly effective way to predict and interpretable in professional network analytics, which has both a methodological and practical implication on understanding digital influence.

2026-01-28T00:00:00Z