Browsing by Subject "statistical science"

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    Open Access
    A review-aware multi-modal neural collaborative filtering recommender system
    (2024) Singh, Pavan; Durbach, Ian; Clark Allan E
    Online shopping has become a ubiquitous aspect of modern life and recommender systems have become a crucial tool for e-commerce giants to efficiently sift through vast amounts of data to locate the infor mation that users are seeking. Within e-commerce, recommender systems aim to provide users with personalised product recommendations based on their preferences and behaviours. They analyse user data, for example their browsing history, purchase history, and ratings to understand their preferences and make recommendations that align with these preferences. They have become fundamental for information retrieval and provide a particularly lucrative landscape for e-commerce platforms, providing suggestions that effectively prune large information spaces so that users are directed toward those items that best meet their needs and preferences. This thesis looks at developing a neural collaborative filtering (NCF) recommender system model which incorporates data from multi-modalities, textual data and explicit ratings data (and review sentiment). The primary objectives of this study are twofold. Firstly, the aim is to create and assess the efficacy of the, relatively new, deep learning-based collaborative filtering approach - NCF - in comparison to other more traditional collaborative filtering models, commonly used. Secondly, the study seeks to investigate the potential impact of incorporating product review text and review text sentiment in improving the accuracy of recommendations. Our model shall be trained and evaluated on the Amazon Product Reviews dataset, which contains millions of user reviews and feedback on thousands of different products across different categories. The metrics used to evaluate the model include predictive accuracy metrics such as mean absolute error, amongst others, as well as top-n evaluation metrics such as recall@n and precision@n. Our methodology is based on a literature analysis and aims to clearly extrapolate on the recent works which have established a framework for NCF. The results of our study show that the NCF model outperforms all the benchmark models in terms of predictive accuracy and top-n evaluation. The results also show that the inclusion of review text in the NCF model improves the predictive accuracy of the model significantly. The results of this study are significant as they demonstrate the potential benefits of incorporating review text into deep learning-based approaches for collaborative filtering for improved rating prediction.
  • No Thumbnail Available
    Item
    Open Access
    Analysis of the effect of course structure and pattern of usage on the efficacy of online/blended courses
    (2024) Oosthuizen, Andries Cornelius; Scott, Leanne
    Using the Sakai Learning Management System (LMS), this dissertation investigates the structure of course sites in blended and online courses at the University of Cape Town. Then, it evaluates the student interactions that this facilitates. The data selected focused on undergraduate courses in 2019. The student interactions with the tool selected for each site are compared to tool categories that indicate a good academic outcome. The analysis was structured to use four popular unsupervised learning algorithms (K-means, PAM, AGNES, and DIANA) on data sets that included the enrolled users and the tools accessible to students. The clValid package method was used to choose the optimal algorithm and cluster sizes. The findings show that most sites used the default tool selection, with almost half the courses adding outside tools and linking in lecture recordings. Sites with less enrolled students were shown to include more diffuse tools, which allow for more creative pedagogy. The majority of student interactions were for course development and delivery, followed by grading and assessment. Finally, most students utilised the LMS and accessed a high percentage of tools in each category. However, the analysis had certain limitations about the events tracked by the system and assumed a one-sided perspective as only the student interaction with the LMS was considered.
  • Thumbnail Image
    Item
    Open Access
    Modelling hepatotoxicity in HIV/TB co-infected patients: extensions of the Cox Proportional Hazards Model
    (2020) Mlotshwa, Vintia Philile; Little, Francesca
    Hepatotoxicity which is also known as liver damage is mainly caused by intake of medicine. It is common among patients who are co-administering Tuberculosis (TB) treatment and the antiretroviral therapy (ART) for the Human Immunodeficiency Viruses (HIV). If severe, hepatotoxicity sometimes necessitates cessation or interruption of treatment. Therefore, understanding, monitoring and managing hepatotoxicity in patients co-infected with TB and HIV is crucial for optimal treatment outcomes. Hepatotoxicity has been investigated in patients coinfected with TB and HIV, however, most studies have analyzed only the first occurrence of hepatotoxicity and discarded information relating to the resolution and recurrence of hepatotoxicity. Data from the ‘Starting Antiretroviral therapy at three Points in Tuberculosis' (SAPiT) trial is used in this project. This was a trial that was instrumental in finalizing treatment guidelines for patients co-infected with HIV and TB in South Africa. The clinical objectives of this project are to estimate incidence rates and determine risk factors associated with hepatotoxicity. The statistical objectives are to fit a Cox regression model, the resolution model of hepatotoxicity, and the extended Cox models for recurring events, including the Andersen Gill (AG) model, the Shared frailty model, the Prentice, Williams and Peterson (PWP) total time (TT) model, the PWP gap time (GT) model, as well as a Cox based recurrent model, that models only the second occurrence of hepatotoxicity. There were 593 patients assessed for hepatotoxicity in the study, 30% (179/593) developed the first occurrence of hepatotoxicity (grade >=1) and 2% (13/593) developed severe hepatotoxicity (grade >=3). Resolved cases (grade = 0) are 76% (136/179) and recurring cases (grade >=1) 24% (32/136). In the Cox multivariable analyses: time-varying treatment arm, older patients, alcohol consumption, low baseline total bilirubin and a positive baseline Hepatitis B surface antigen status, were associated with a higher risk of developing the first occurrence of hepatotoxicity. The extended Cox models (AG model, Shared frailty model, PWP TT model and PWP GT model) in combination identified that: time-varying treatment arm, older patients, alcohol consumption, baseline CD4 count that is greater than 50 cells per mm3 , low baseline total bilirubin, and a positive baseline Hepatitis B surface antigen status were associated with an increased risk of developing recurrent hepatotoxicity. In the resolution model multivariable analyses; non-consumers of alcohol and an abnormal liver function tests at baseline, were associated with an increased chance of resolving the first occurrence of hepatotoxicity. In the multivariable analyses of the recurrent model: younger patients and the time-varying treatment arm were associated with the development of the second occurrence hepatotoxicity. Since the Cox regression model utilized data up to the first occurrence of hepatotoxicity, in some instances, the time-varying treatment effect based on the Cox regression model was closer to unity and marginally significant. And the corresponding effect based on the recurrent event models (AG model, Shared frailty model, PWP TT model, PWP GT model and the recurrent model), that utilized data of the first and second occurrence of hepatotoxicity, generally produced a time-varying treatment effect slightly far from unity with a strong statistical significance. This trend was similar for other predictors of hepatotoxicity, like CD4 count and alcohol consumption. In conclusion, hepatotoxicity is common in this study, however, it is often transient or mild and did not necessitate treatment interruption. However, close monitoring of patients especially in the first 5 months of TB-treatment is recommended. The PWP TT model seemed to be the best model for modelling recurring hepatotoxicity, since the identified risk factors that were associated with hepatotoxicity, changed from the first occurrence of hepatotoxicity to the second occurrence of hepatotoxicity.
  • Thumbnail Image
    Item
    Open Access
    Resource constraints in an epidemic: a goal programming and mathematical modelling framework for optimal resource shifting in South Africa
    (2021) Mayet, Saadiyah; Silal, Sheetal; Durbach, Ian
    The COVID-19 pandemic has had devastating consequences across the globe, and has led many governments into completely new decision making territory. Developing models which are capable of producing realistic projections of disease spread under extreme uncertainty has been paramount for supporting decision making by many levels of government. In South Africa, this role has been fulfilled by the South African COVID-19 Modelling Consortium's generalised Susceptible-ExposedInfectious-Removed compartmental model, known as the National COVID-19 Epi Model. This thesis adapted and contributed to the Model; its primary contribution has been to incorporate the feature that resources available to the health system are limited. Building capacity constraints into the Model allowed it to be used in the resource-scarce context of a pandemic. This thesis further designed and implemented a goal programming framework to shift ICU beds between districts intra-provincially in a way that aimed to minimise deaths caused by the non-availability of ICU beds. The results showed a 15% to 99% decrease in lives lost when ICU beds were shifted, depending on the scenario considered. Although there are limitations to the scope and assumptions of this thesis, it demonstrates that it is possible to combine mathematical modelling with optimisation in a way that may save lives through optimal resource allocation.