Master Theses: Department of Computing and Information Technology

Permanent URI for this collectionhttp://repository.embuni.ac.ke/handle/embuni/3869

Browse

Now showing 1 - 3 of 3

Sentiment Analysis-Based Model for Monitoring User Engagement With Mental Health Chatbots
(UoEm, 2025-10-01) MMBAYI, IAN IGADO
Mental health challenges, particularly among youth, are compounded by stigma and limited access to professional care. This has driven demand for scalable digital solutions like chatbots. This study introduces a sentiment analysis-based model to assess user satisfaction with mental health chatbots, analysing 82,102 reviews from six popular apps on Google Play and Apple’s App Stores. A multi-class sentiment classification of positive, negative, and neutral was implemented, enhanced by Synthetic Minority Over-sampling Technique for class balancing, comparing five traditional machine learning models with Bidirectional Encoder Representations from Transformers, a transformer model. Random Forest achieved 98.18% accuracy among traditional models, while BERT outperformed all with 99.17% accuracy, surpassing prior benchmarks. Aspect-based analysis revealed that Emotion and Usability drive positive feedback, while Reliability issues fuel negative sentiments, offering actionable insights for developers to enhance chatbot design. This work advances digital mental health research by integrating multi-class classification, transformer models, and aspect-based analysis, demonstrating a scalable framework for evaluating user feedback.
Implementation of A Hybrid Model Using K-Means Clustering and Artificial Neural Networks for Risk Prediction in Life Insurance
(UoEm, 2024-09) Kimanga Nthenge, Jeff
Accurate assessment of the risk posed by prospective policyholders is crucial for life insurance companies to effectively price policies and manage long-term liabilities. However, the complexity of risk factors makes relying solely on traditional actuarial models insufficient, particularly with the abundance of big data and unstandardized data from various sources. This study explored the development and performance of a hybrid machine learning model that combines Artificial Neural Network and K-Means Clustering to improve risk prediction in life insurance underwriting. A quasi-experimental design was adopted to evaluate the efficacy of K-Means Clustering and ANN algorithms on benchmark datasets and develop a hybrid model for risk prediction. The proposed hybrid model utilized the strengths of Artificial Neural Networks in modelling nonlinear relationships and K-Means in pattern recognition to handle unstandardized data. Using anonymized life insurance application data from Kaggle, the ANN algorithm achieved an accuracy of 90% but showed limitations in handling nonlinear relationships. K-Means Clustering successfully identified distinct risk profiles among policyholders, revealing hidden patterns in the unlabelled data. The hybrid model, integrating K-Means Clustering and ANN with principal component analysis for feature selection and the Adam optimizer, resulted in higher model performance. Testing accuracy improved from 90% for the standalone ANN to 98% for the hybrid technique, with improvements in precision, recall, and Area Under the ROC Curve. The enhanced predictive capability highlighted the potential of the hybrid approach in modernizing underwriting practices and conducting a more sophisticated data-driven analytical evaluation of policyholder risk. However, there were limitations, such as the use of a single-sourced insurance dataset due to concerns about data privacy. Further research into integrating diverse algorithms and testing on larger real-world datasets can assist insurers in unlocking more value and gaining a competitive advantage through advanced analytical modelling.
Protecting Institutions of Higher Learning in Kenya: A Scalable Hybrid Decoy Framework against Cyber Threats
(University of Embu, 2021-09) Serem, Edwin Kiprono
Cybersecurity threats are malicious acts that seek to damage, steal, or gain unauthorized access to information. Higher institutions of learning in Kenya have adopted the use of information systems in their service delivery. However, their level of preparedness to deal with emerging threats in their cyberspace is limited by techniques used to detect, inform, and deflect the cyber threats before they cause much harm. The main objective of this research study was to develop a scalable decoy framework for use in institutions of higher learning. The research process was done in two phases; the first phase encompassed preliminary studies that involved soliciting responses from 84 ICT personnel drawn from 42 institutions in Kenya selected through the purposive sampling method. This study made use of primary data collected using structured questionnaires, then descriptively analyzed. The findings revealed the institutions recorded cyber attacks within twelve months of the research period, and the main tools and techniques in place are inefficient to detect significant threats. The second phase entailed designing the framework prototype using Linux containers as decoys in the front and back end and monitoring the attacks using HonSSH, while graphical presentation used Grafana. The decoys were set in a layered approach. The front-end decoy hid the back-end decoy by internally configuring the front-end decoy to capture and reroute the attacker commands via a secure tunnel. The back-end decoy did the processing of commands issued through the front-end decoy then gave feedback. Simulation of user activities and network traffic generation was achieved using the General HOSTS framework to make it more realistic to the attacker. The attacker's virtual machine used Kali Linux. Scalability, latency, and throughput metrics were used to test the framework's effectiveness; decoy data analysis was done by logstash and pipelined to Kibana for visualization. The experimental results demonstrate that the system effectively misdirected commands by combining deceptive network setup and configurations and generating fake user and network activities with an average latency of 0.0015s, throughput 864Mbits/s, and boot speed 7.485s. The study highly recommends including cyber decoys in the institutions network to boost security in a proactive approach due to effectiveness in utilizing computing resources. The framework will help cybersecurity professionals protect higher institutions of learning from stealthy and sophisticated attacks. This research work contributes to knowledge in designing and developing effective deceptive decoys tools in cybersecurity research.

Browse

Recent Submissions