Pusat Pengajian Sains Komputer - Tesis
Browse
Recent Submissions
Now showing 1 - 5 of 472
- PublicationA Fusion-Based Framework For Explainable Suicide Attempt Prediction(2024-05)Nordin, NoratikahSuicide remains a major public health problem and one of the leading causes of death worldwide. Suicide prevention is needed to reduce global suicide mortality, as highlighted in the united nations third sustainable development goals (sdgs). A suicide attempt is the most complex and dynamic suicidal behaviour, which is important for suicide prevention strategies. However, decision-making in classifying individuals at higher risk of suicide attempts is subjective and uncertain. Existing studies on the framework for predictive models using data-driven and knowledge-driven approaches are insufficiently explained and unable to provide an understandable prediction of suicide attempts for suicide prevention in a systematic way. Therefore, this study presents a fusion-based framework for explainable suicide attempt prediction using explainable data-driven and knowledge-driven approaches to classify and explain individuals with suicide attempts to support decision-making by medical experts. The proposed work aims to analyse an explainable learning algorithms for predicting suicide attempts, propose an ontology model for semantically representing the classification risk of suicide attempts and propose an explanation generation algorithm by combining predictions from explainable machine learning and ontology models. An information fusion-based explanation generation method is proposed by integrating predictions to generate a prediction description to support decision-making. The fusion model shows that the proposed framework achieves 92% accuracy, 88% specificity, and 100% sensitivity.
- PublicationMultimodal Sentiment Analysis Of Social Media Through Deep Learning Approach(2020-06)An, JieyuMultimodal Data, Characterized By Its Inherent Complexity And Heterogeneity, Presents Computational Challenges In Comprehending Social Media Content. Conventional Approaches To Sentiment Analysis Often Rely On Unimodal Pre-Trained Models For Feature Extraction From Each Modality, Neglecting The Intrinsic Connections Of Semantic Information Between Modalities, As They Are Typically Trained On Unimodal Data. Additionally, Existing Multimodal Sentiment Analysis Methods Primarily Focus On Acquiring Image Representations While Disregarding The Rich Semantic Information Contained Within The Images. Furthermore, Current Methods Often Overlook The Significance Of Color Information, Which Provides Valuable Insights And Significantly Influences Sentiment Classification. Addressing These Gaps, This Thesis Explores Deep Learning-Based Methods For Multimodal Sentiment Analysis, Emphasizing The Semantic Association Between Multimodal Data, Information Interaction, And Color Sentiment Modelling From The Perspectives Of The Multimodal Representation Layer, The Multimodal Interaction Layer, And The Color Information Integration Layer. To Mitigate The Overlooked Semantic Interrelations Between Modalities, The Thesis Introduces "Joint Representation Learning For Multimodal Sentiment Analysis" Within The Representation Layer. This Method, Validated By Rigorous Experiments, Showcases A Marked Improvement In Accuracy, Achieving 76.44% On The Mvsa-Single And 72.29% On The Mvsa-Multiple Datasets, Surpassing Existing Methodologies. In The Multimodal Interaction Layer,
- PublicationMental Stress Classification Among Higher Education Students In Malaysia From Electroencephalogram (Eeg) Using Convolutional Neural Network With Modified Stochastic Gradient Descent(2024-09)Rashid, Nur Ramizah RaminoThis study investigates the classification of mental stress among Malaysian university students using Electroencephalogram (EEG) data and a 1D-Convolutional Neural Network (1D-CNN) optimized with Modified Stochastic Gradient Descent (SGD). The research addresses a significant gap in the availability of localized datasets for stress detection using EEG signals, as existing models and datasets predominantly focus on other populations and do not account for regional variations in stressors and responses. Moreover, there is a lack of optimization in stress detection models, specifically in handling EEG data, which can affect the models’ accuracy and real-time application potential. To address these challenges, EEG signals were collected during Stroop tests and self-reported stress levels were measured using the Perceived Stress Scale (PSS). A rigorous preprocessing approach, including Independent Component Analysis (ICA) for artifact removal, was applied, followed by feature extraction focusing on key metrics such as energy, entropy, and standard deviation from both time and frequency domains. The chosen algorithm, 1D-CNN, was modified using a tailored SGD optimizer that incorporates momentum and learning rate decay to improve convergence and address challenges like vanishing gradients. This modification was essential for enhancing the model’s learning process, ultimately leading to better stress classification performance. The proposed 1D CNN model, enhanced with Modified SGD, demonstrated superior performance compared to traditional models such as Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and deeper architectures like Standard CNN and AlexNet. Specifically, the 1D CNN achieved an accuracy of 92.64%, outperforming SVM (84.5%), k-NN (76.6%), Standard CNN (91.3%), RNN (90.04%) and AlexNet (91.65%). The 1D CNN model also demonstrated high sensitivity and specificity, making it a robust solution for EEG-based stress detection.
- PublicationModified Nsga-Iii As A Many-Objective Optimization Technique For Ab Initio Protein Structure Prediction(2024-05)Smadi, Motasem Mohammad Yaqoob AlPredicting the 3D structure of proteins, known as protein structure prediction (PSP), is an essential component of bioinformatics. One of the adopted approaches to solving the PSP problem is an ab initio modeling method, which has demonstrated great performance in recent years. In this work of research, the non-dominated sorting genetic algorithm III (NSGA-III) has been adapted by modeling an ab initio PSP problem as a many-objective optimization problem (MOOP). Furthermore, this study proposes four novel objectives from three distinct energy function types with the goal of solving as a MOOP. In this regard, the first two objectives are derived from the physics-based energy function (PEF) which decomposes into bond and nonbond terms. Besides, the third objective is a solvent effect model represented by SASA, and the last objective is a knowledge-based energy function (KEF) represented by dDFIRE. To demonstrate the efficiency of energy functions, the implemented NSGA-III is handled with three suggested sets of objectives: MO3SASA (consisting of bond, nonbond, and SASA); MO3dDFIRE (consisting of bond, nonbond, and dDFIRE); and MaO4 (consisting of bond, nonbond, SASA, and dDFIRE). Moreover, to further improve the exploration and exploitation of conformational space search processes with regard to NSGA-III, three genetic operators of the genetic algorithm (GA) are proposed, including two selection operators, four crossover operators, and three mutation operators. These proposed operators produced 25 conformational search algorithms (CSAs) for the proposed NSGA-III.
- PublicationSoftware Model Checking For Distributed Applications Using Hybridization Of Centralization And Cache Approaches(2024-04)Hing, RatanaDeveloping reliable distributed systems poses significant challenges due to the non-deterministic nature of thread and process execution, as well as communication channels. Software model checking offers a means to verify system correctness by exhaustively analyzing all program execution paths. However, the existing bytecode model checker, capable of verifying multiple processes, suffers from state space explosion and computational overhead. This thesis introduces Java PathFinder (JPF)-Nas-Hybrid (JNH), a novel model checker addressing these limitations. JNH employs a redesigned inter-process communication (IPC) model and integrates a scalable caching mechanism. This mechanism efficiently stores communication data between processes, mitigating computational overhead and state space explosion during model checking. By optimizing resource utilization and minimizing overhead, JNH significantly improves verification performance. Key enhancements include the development of a scalable caching mechanism integrated into the centralization IPC model, relocating request and response trees, and processing data in multi-byte chunks. JNH's creation involves extending from the JPF-core system and modifying Java network libraries. Additionally, the thesis explores bug detection strategies, distinguishing between local and global bugs, and evaluates various search strategies to explore distributed program state spaces. Through comprehensive testing and statistical analysis, the research provides insights into effective bug detection approaches, further advancing model-checking methodologies.