Institutional-Repository, University of Moratuwa
Welcome to the University of Moratuwa Digital Repository, which houses postgraduate theses and dissertations, research articles presented at conferences by faculties and departments, university-published journal articles and research publications authored by academic staff. This online repository stores, preserves and distributes the University's scholarly work. This service allows University members to share their research with a larger audience.
 Research Publications |  Thesis & Dissertation |  E- Books |
 UoM Journal Publications |  UoM Conference Proceedings |  Articles published in Scimago's Q1 journals |  UoM Research Reports |  Other Articles authored by UoM staff |
Recent Submissions
item: Thesis-Abstract
Enhancing employee retention : an improved hybrid model for predicting employee attrition
(2025) Madhubhashini, JMM; De Silva, CR
Employee attrition poses a significant challenge to organizational stability, directly impacting productivity, operational continuity, and long-term strategic goals. While traditional methods of attrition management rely on retrospective analyses, machine learning (ML) offers a proactive approach to predict and mitigate workforce turnover. This study addresses the critical need for robust predictive frameworks by evaluating eight ML classifiers; Support Vector Machines with Radial Basis Function kernel, Fisher’s Linear Discriminant, Logistic Regression, Multi-Layer Perceptron, Random Forests, Naive Bayes, Adaboost and XGBoost with IBM HR Analytics Employee Attrition & Performance dataset to forecast employee attrition. A novel hybrid stacking ensemble model is proposed to enhance prediction accuracy by integrating the strengths of individual classifiers. The research investigates the impact of dataset balancing and feature optimization on model performance, emphasizing the role of data preprocessing in improving predictive reliability. Models underwent hyperparameter tuning and stratified 5-fold cross-validation, with the hybrid ensemble (Logistic Regression meta-learner) achieving superior metrics: 95.81% accuracy, 96.69% Precision, 95.77% F1-score, and 99.23% ROC-AUC. Individual models exhibited strong performance, notably XGBoost and LR (93.38% accuracy), while dataset balancing improved minority-class recall by 18–22% and feature selection reduced redundancy by 30%. Key findings reveal that dataset balancing and feature engineering significantly enhance model precision, particularly for minority class identification. The stacking model’s superior performance underscores the value of meta-learning in synthesizing heterogeneous classifier outputs. Practical implications include scalable frameworks for HR analytics, enabling early risk identification and personalized retention strategies. The stacking model’s adaptability allows integration with HR platforms, transforming reactive practices into proactive decision-making. Methodologically, the study advances predictive human capital management by demonstrating the synergy of hybrid ML approach and interpretable outputs. By bridging technical innovation with organizational psychology, this work offers a replicable blueprint for sustainable workforce management, emphasizing real-time analytics and employee-centric interventions.
item: Thesis-Abstract
Identifying emotional valence from FMRI data
(2025) Edirisinghe, KEABI; Chithraranjan, C
Identifying emotional valence from fMRI data is a significant step toward understand- ing how the brain processes affective states. This has important implications in fields such as cognitive neuroscience, mental health, and affective computing. However, decoding emotional valence using machine learning remains a challenging task, espe- cially due to the high dimensionality of fMRI data, inter-subject variability, and the subtle nature of emotional responses. Achieving high classification accuracy is partic- ularly difficult, making it crucial to explore effective feature extraction methods. This study focuses on comparing the efficiencyof two popular featureengineering approaches— General Linear Model (GLM) -based methods and Independent Com- ponent Analysis (ICA) —for emotion classification using fMRI data. Specifically, we implement the GLMSingle method for GLM-based feature extraction and compare it against ICA-derived features. The extracted features are evaluated using multiple ma- chine learning classifiers, including Support Vector (SVM) , Gaussian Naive Bayes (GNB), RandomForest(RF),and LogisticRegression(LR).Weassessclassification performance both within subjects and across subjects to evaluate generalizability. The results show that features derived from the GLMSingle method consistently outperform those from ICA in terms of classification accuracy. This suggests that task-related modeling using GLM provides more discriminative features for emotion prediction than the data-driven ICA approach. The findings highlight the importance of choosing suitable feature engineering strategies in order to improve model perfor- mance when decoding affective states from neuroimaging data.
item: Conference-Full-text
A Novel mechanism or improving safety and productivity of coconut deshelling
(IEEE, 2024) Bandara, A; Jayawickrama, D; Kuladasa, B; Lihil Uthpala Subasinghe, LU; Gamage, JR
Coconut based products contributes 6.97% of total merchandise exports of Sri Lanka in 2024 [1]. Deshelling is as a key process step in the manufacturing of coconut-based food products. However, industrial deshelling methods poses serious safety risks, damages to the kernel, and high dependency on skilled labour. This paper presents a novel coconut deshelling approach to address the challenges identified in the current process. A review of literature was used to identify the state-ofthe-art of current deshelling methods and to establish the critical parameters involved in deshelling. Morphological analysis was conducted to develop conceptual designs. Industrial visits were conducted to get the real need of the industry to understand the current process. The design concepts were benchmarked against the parameters identified through the review and industrial visits. Finalised design concept is modelled using SolidWorks® and a prototype was fabricated. The prototype was improved with several iterations and experimented for intended performance. The prototype machine consists of two stations, cutting station and shell removing station. The proposed machine does not require skilled labour thus one operator can oversee several machines simultaneously. The safety threat to the operator is eliminated in the proposed machine and current productivity can be met with multiple machines until the machine gets adequate physical realisation of the design with sufficient use of resources.
item: Thesis-Full-text
Explainable AI for breast cancer detection in mammography
(2025) Wickremesinghe, LLM; Thanuja, DA
Breast cancer remains a significant global health concern among women. This research introduces an explainable AI-assisted breast cancer detection system aimed at improving both the accuracy and interpretability of mammogram-based diagnoses. The study utilizes high-quality mammographic datasets, CBIS-DDSM and the RSNA Screening Mammography dataset, to train and validate the models. The system uses two powerful deep learning models: Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). The InceptionResNetV2 CNN achieved an accuracy of 92%, while the ViT model reached 96% accuracy by effectively focusing on important regions in the mammogram images. To make the system more transparent, several Explainable AI (XAI) methods were applied, including Grad-CAM, SIDU, Attention Maps, and Ablation-CAM. Among these, SIDU provided the clearest and most accurate visual explanations, which are valuable for medical decisionmaking. To further improve the reliability and clinical value of the system, this study introduces a Dual-Stage Ensemble Diagnosis and Decision Fusion Framework. This approach combines the diagnostic strengths of both models to deliver a more confident and balanced final decision, supported by detailed visual explanations. The platform consist with a user-friendly web application that allows doctors and patients to easily upload mammogram images and receive AI-based predictions with clear and interpretable outputs. This research helps advance the development of trustworthy AI tools for breast cancer detection in real clinical settings.
item: Conference-Full-text
Assessing the prevalence of energy poverty: a field investigation of households with elders in underserved communities in Colombo
(IEEE, 2024) Chathuranga, N; Rajapaksha, I; Sajjad, My; Siriwardana, C
Energy is a primary resource for a nation's socioeconomic development. Targeted policy interventions to ensure energy equity and alleviate energy poverty must be grounded in a comprehensive understanding of the prevailing issues in energy accessibility and affordability among vulnerable communities. This study focuses on assessing the incidence and intensity of energy poverty among households inhabited by elders in underserved communities in Colombo, Sri Lanka. The study assesses prevailing energy poverty using the Multidimensional Energy Poverty Index (MEPI) across 149 randomly selected households. Various indicator weighting methods and energy poverty cut-off values are utilized in MEPI calculations for comparative analysis. These households experience an average MEPI of 0.262 with a 0.476 headcount ratio. Additionally, 77.4% of households spend over 10% of their income on energy services, indicating high energy expenditure relative to household income. Moreover, most of these households face numerous challenges, including confined floor areas, temporary construction, inadequate natural lighting, poor ventilation and air quality, and inadequate water and sanitation facilities. Thus, prioritizing the needs of elders in these communities and implementing evidence-based policy interventions are crucial steps toward fostering a more equitable and inclusive society. The study's outcomes have farreaching implications for strategic interventions aimed at alleviating energy poverty.