4. Electronic Theses and Dissertations (ETDs) - Faculties submissions

Permanent URI for this communityhttps://hdl.handle.net/10539/37773

For queries relating to content and technical issues, please contact IR specialists via this email address : openscholarship.library@wits.ac.za, Tel: 011 717 4652 or 011 717 1954

Browse

Search Results

Now showing 1 - 10 of 15

Prediction of Water Hyacinth Coverage on Hartbeespoort Dam
(University of the Witwatersrand, Johannesburg, 2024) de Gouveia, Claudia D. Camacho; Bührmann, Doctor Joke
Water hyacinth is an invasive weed contributing to Hartbeespoort Dam’s poor water quality. Although biological control is the most effective and sustainable method of controlling water hyacinth, the dam has unfavourable conditions for agents that the weed thrives in. Literature uses mathematical models and remote sensing to theorise growth rates or estimate coverage. However, prediction could prove beneficial as planning biological control is essential to its success. Hence, a model to predict water hyacinth coverage was developed. This research simplified the complex relationships involved in water hyacinth growth to focus on the most influential factors: temperature and nutrients. Missing data were imputed using multiple k-nearest neighbours. Nutrient datasets had limited data, thus five scenarios were developed to extrapolate datasets, using Monte Carlo simulation and seasonal patterns. The features were used to build ensemble, decision tree, artificial neural network and support vector machine models. Ensemble using the bagging method was the best model resulting in a root mean square error of 4.01 for water hyacinth coverage predictions from 1 June 2018 to 1 May 2019.
Machine Learning Algorithms-Based Classification of Lithology using Geophysical Logs: ICDP DSeis Project Boreholes, South Africa
(University of the Witwatersrand, Johannesburg, 2024-09) Atita, Obehi Chapet; Durrheim, Raymond; Saffou, Eric
One of the most significant geosciences tasks is the accurate classification of lithologies for metal and mineral resources exploration, characterization of oil/gas reservoir(s), and the planning and management of mining operations. With the availability of abundant, huge and multidimensional datasets, machine learning-based data-driven methods have been widely adopted to assist in solving geoscientific problems such as the efficient evaluation and interpretation of large datasets. The adoption of machine learning-based methods aims to improve lithological identification accuracy and extract information required for accurate and objective decision-making with respect to activities such as exploration, drilling, mine planning and production. Practically, this helps to reduce working time and operating costs. We aim to evaluate the feasibility of machine learning-based algorithms application to geophysical log data for the automated classification of lithologies based on the stratigraphic unit at the formation level for the purpose of distinguishing and correlating the quartzites between boreholes, and mapping key radioactive zones within the mining horizon. This study implemented four different machine learning algorithms: gradient boosting decision trees, random forest, support vector machine, and K-means clustering models. Analyzed features and labelled datasets are multivariate downhole geophysical and lithology logs from the two ICDP DSeis project boreholes drilled in the Klerksdorp gold field, respectively. To mitigate misclassification error and avoid model overfitting/underfitting, the optimal combination sets and optimal values for each implemented supervised model’s hyperparameters were obtained using the Grid search and 10-fold cross-validation optimization methods. The input dataset was randomly split automatedly into training and testing subsets that made up 80% and 20% of the original dataset, respectively. The models were trained and cross-validated using the training subset, and their performances were assessed using the testing subset. The classification performance of each model was evaluated using F1 scores and visualized using confusion matrices. The best supervised classification model for our study area was selected based on the testing subset F1 scores and computational cost of training models. The testing subset results shows that Random Forest and Support Vector Machine classifier models performed much better relative to the Gradient Boosting Decision Trees classifier model, with F1 scores over 0.80 in borehole A and B. In borehole A and B, Random Forest classifier has the least computational training time of about 14- and 6- hours, respectively. The feature importance results demonstrate that the logging feature P-wave velocity (Vp) is the highest predicting feature to the lithology classification in both boreholes. We find that the quartzite classes at different stratigraphic positions in each borehole are similar and they are correlated between the DSeis boreholes. The K-means clustering revealed three clusters in this study area and effectively map the radioactive zones. This study illustrates that geophysical log data and machine learning-based algorithms can improve the task of data analysis in the geosciences with accurate, reproducible and automated prediction of lithologies, correlation and mapping of radioactive zones in gold mine. This study outputs can serve as quality control measures for future similar studies both in the academic and industry. We identified that availability of large data is the major factor to high accuracy performance of machine learning-based algorithms for classification problems.
BiCoRec: Bias-Mitigated Context-Aware Sequential Recommendation Model
(University of the Witwatersrand, Johannesburg, 2024-09) Muthivhi, Mufhumudzi; van Zyl, Terence; Bau, Hairong
Sequential recommendation models aim to learn from users’ evolving preferences. However, current state-of-the-art models suffer from an inherent popularity bias. This study developed a novel framework, BiCoRec, that adaptively accommodates users’ changing preferences for popular and niche items. Our approach leverages a co-attention mechanism to obtain a popularity-weighted user sequence representation, facilitating more accurate predictions. We then present a new training scheme that learns from future preferences using a consistency loss function. The analysis of the experimental results shows that our approach is 7% more capable of uncovering the most relevant items.
Assessing and comparing the performance of different machine learning regression algorithms in predicting Chlorophyll-a concentration in the Vaal Dam, Gauteng
(University of the Witwatersrand, Johannesburg, 2024-03) Mahamuza, Phemelo Hope; Adam, Elhadi
The state of Vaal Dam is influenced by various land uses surrounding the Dam, including agricultural activities, mining operations, industrial enterprises, urban settlements, and nature reserves. Mining activities, farming practices, and sewage outflows from nearby villages led to access contamination within the Dam, increasing algal bloom levels. Sentinel-2 MSI data were utilized to forecast and comprehend the spatial pattern of Chlorophyll-a concentration, indicating algal bloom occurrence in the Vaal Dam. Targeting Sentinel-2 Level-1C, the image was preprocessed on the Google Earth Engine (GEE) with acquisition dates from 25 – 26 October 30, 2016, corresponding to the on-site data collection between October 26 and October 28, 2016. Due to limited resources, up-to-date data on the Vaal Dam could not be collected. However, since this study focuses on applying various machine learning regression models to predict chlorophyll-a levels in waterbodies, the dataset is used to test the models rather than reflect the current state of the Vaal Dam. The dataset, comprising 23 samples, was divided into 70% training and 30% test sets, allowing for comprehensive model evaluation. Band ratio reflectance values were extracted from the satellite image and correlated with in-field Chlorophyll-a values. The highest correlation coefficient values were utilized to train five machine-learning models employed in this study: Random Forest (RF), Support Vector Regression (SVR), Least Absolute Shrinkage and Selection Operator (LASSO), Ridge Regression, and Multilinear Regression (MLR). Each model underwent training with ten iterations each; the best learning iteration was then used to generate the final Chlorophyll-a predictive model. The predictive models were validated using the Sentinel-2 MSI satellite data and in-situ measurements using R2, RMSE, and MAPE. Among the five machine learning algorithms trained, RF performed the best, with an R2 of 0.86 and 0.95, an RMSE of 1.38 and 0.8, and MAPE of 15.09% and 10.92% for the training and testing sets, respectively, indicating its ability to handle small, non-linear datasets. SVR also demonstrated a fair performance, particularly in handling multicollinearity in the data points with an R2 of 0.68 and 0.87, an RMSE of 2.37 and 1.56, and MAPE of 18.13% and 19.28% for the training and testing sets, respectively. The spatial pattern of Chlorophyll-a concentrations, mapped from the RF model, indicated that high concentrations of Chlorophyll-a are along the Dam shorelines, suggesting a significant impact of land use activities on pollution levels. This study emphasizes the importance of selecting suitable machine learning algorithms tailored to the dataset's characteristics. RF and SVR demonstrated proficiency in handling nonlinearity, with RF displaying enhanced generalization and resistance to overfitting. Limited field data evenly distributed across the Dam and satellite overpass dates may affect result accuracy. Future research should align satellite pass dates with fieldwork dates and ensure an even distribution of in-field samples across the Dam to represent all land uses and concentration levels.
3D Human pose estimation using geometric self-supervision with temporal methods
(University of the Witwatersrand, Johannesburg, 2024-09) Bau, Nandi; Klein, Richard
This dissertation explores the enhancement of 3D human pose estimation (HPE) through self-supervised learning methods that reduce reliance on heavily annotated datasets. Recognising the limitations of data acquired in controlled lab settings, the research investigates the potential of geometric self-supervision combined with temporal information to improve model performance in real-world scenarios. A Temporal Dilated Convolutional Network (TDCN) model, employing Kalman filter post-processing, is proposed and evaluated on both ground-truth and in-the-wild data from the Human3.6M dataset. The results demonstrate a competitive Mean Per Joint Position Error (MPJPE) of 62.09mm on unseen data, indicating a promising direction for self-supervised learning in 3D HPE and suggesting a viable pathway towards reducing the gap with fully supervised methods. This study underscores the value of self-supervised temporal dynamics in advancing pose estimation techniques, potentially making them more accessible and broadly applicable in real-world applications.
Predicting in-hospital mortality in heart failure patients using machine learning
(University of the Witwatersrand, Johannesburg, 2023-05) Mpanya, Dineo; Ntsinjana, Hopewell
The age of onset and causes of heart failure differ between high-income and low-and-middle-income countries (LMIC). Heart failure patients in LMIC also experience a higher mortality rate. Innovative ways that can risk stratify heart failure patients in this region are needed. The aim of this study was to demonstrate the utility of machine learning in predicting all-cause mortality in heart failure patients hospitalised in a tertiary academic centre. Six supervised machine learning algorithms were trained to predict in-hospital all-cause mortality using data from 500 consecutive heart failure patients with a left ventricular ejection fraction (LVEF) less than 50%. The mean age was 55.2 ± 16.8 years. There were 271 (54.2%) males, and the mean LVEF was 29 ± 9.2%. The median duration of hospitalisation was 7 days (interquartile range: 4–11), and it did not differ between patients discharged alive and those who died. After a prediction window of 4 years (interquartile range: 2–6), 84 (16.8%) patients died before discharge from the hospital. The area under the receiver operating characteristic curve was 0.82, 0.78, 0.77, 0.76, 0.75, and 0.62 for random forest, logistic regression, support vector machines (SVM), extreme gradient boosting, multilayer perceptron (MLP), and decision trees, and the accuracy during the test phase was 88, 87, 86, 82, 78, and 76% for random forest, MLP, SVM, extreme gradient boosting, decision trees, and logistic regression. The support vector machines were the best performing algorithm, and furosemide, beta-blockers, spironolactone, early diastolic murmur, and a parasternal heave had a positive coefficient with the target feature, whereas coronary artery disease, potassium, oedema grade, ischaemic cardiomyopathy, and right bundle branch block on electrocardiogram had negative coefficients. Despite a small sample size, supervised machine learning algorithms successfully predicted all-cause mortality with modest accuracy. The SVM model will be externally validated using data from multiple cardiology centres in South Africa before developing a uniquely African risk prediction tool that can potentially transform heart failure management through precision medicine.
Machine learning in marketing strategy: A socio-technical approach in South Africa
(University of the Witwatersrand, Johannesburg, 2024) Govender, Aleasha; Quaye, Emmanuel
The purpose of this research study was to determine whether the existing market segmentation, targeting and positioning (STP) approaches are optimal for marketing strategy in South Africa, and to what extent AI and machine learning are being used to improve marketing strategy in South Africa. The methods used have drawn on qualitative data research and document analysis. There were 10 participants in the study, the industries include Banking, Telecommunication and Medical Insurance. The methods used have drawn on qualitative data research and document analysis. The key results of the research have determined that Machine Learning is in its inception phase in terms of being used in marketing strategy in corporate South Africa. The research further finds that there are factors that are slowing the development in this field that are aligned with both hard and soft capabilities, for example, along with infrastructural capabilities like software integration, strategic capabilities like interdepartmental alignment are required for effective deployment of these technologies. Further, the research finds that the current segmentation, targeting and positioning methods used in isolation are not optimally contributing to marketing strategy, rather a blended approach including insights from customer data will provide a more accurate STP strategy. This research supports marketeers, technologists, business structures, researchers in South Africa, as well as strategists who deal with mass consumer bases, because market segmentation, targeting and positioning underpin how marketing strategy is rolled out throughout corporate South Africa and AI and Machine Learning are emerging technologies that are highly topical and are only at the inception phase of optimal utilisation
Generating Rich Image Descriptions from Localized Attention
(University of the Witwatersrand, Johannesburg, 2023-08) Poulton, David; Klein, Richard
The field of image captioning is constantly growing with swathes of new methodologies, performance leaps, datasets, and challenges. One new challenge is the task of long-text image description. While the vast majority of research has focused on short captions for images with only short phrases or sentences, new research and the recently released Localized Narratives dataset have pushed this to rich, paragraph length descriptions. In this work we perform additional research to grow the sub-field of long-text image descriptions and determine the viability of our new methods. We experiment with a variety of progressively more complex LSTM and Transformer-based approaches, utilising human-generated localised attention traces and image data to generate suitable captions, and evaluate these methods on a suite of common language evaluation metrics. We find that LSTM-based approaches are not well suited to the task, and under-perform Transformer-based implementations on our metric suite while also proving substantially more demanding to train. On the other hand, we find that our Transformer-based methods are well capable of generating captions with rich focus over all regions of the image and in a grammatically sound manner, with our most complex model outperforming existing approaches on our metric suite.
Analyzing the performance and generalisability of incorporating SimCLR into Proximal Policy Optimization in procedurally generated environments
(University of the Witwatersrand, Johannesburg, 2024) Gilbert, Nikhil; Rosman, Benjamin
Multiple approaches to state representation learning have been shown to improve the performance of reinforcement learning agents substantially. When used in reinforcement learning, a known challenge in state representation learning is enabling an agent to represent environment states with similar characteristics in a manner that would allow said agent to comprehend it as such. We propose a novel algorithm that combines contrastive learning with reinforcement learning so that agents learn to group states by common physical characteristics and action preferences during training. We subsequently generalise these learnings to previously encountered environment obstacles. To enable a reinforcement learning agent to use contrastive learning within its environment interaction loop, we propose a state representation learning model that employs contrastive learning to group states using observations coupled with the action the agent chose within its current state. Our approach uses a combination of two algorithms that we augment to demonstrate the effectiveness of combining contrastive learning with reinforcement learning. The state representation model for contrastive learning is a Simple Framework for Contrastive Learning of Visual Representations (SimCLR) by Chen et al. [2020], which we amend to include action values from the chosen reinforcement learning environment. The policy gradient algorithm (PPO) is our chosen reinforcement learning approach for policy learning, which we combine with SimCLR to form our novel algorithm, Action Contrastive Policy Optimization (ACPO). When combining these augmented algorithms for contrastive reinforcement learning, our results show significant improvement in training performance and generalisation to unseen environment obstacles of similar structure (physical layout of interactive objects) and mechanics (the rules of physics and transition probabilities).
Learning to adapt: domain adaptation with cycle-consistent generative adversarial networks
(University of the Witwatersrand, Johannesburg, 2023) Burke, Pierce William; Klein, Richard
Domain adaptation is a critical part of modern-day machine learning as many practitioners do not have the means to collect and label all the data they require reliably. Instead, they often turn to large online datasets to meet their data needs. However, this can often lead to a mismatch between the online dataset and the data they will encounter in their own problem. This is known as domain shift and plagues many different avenues of machine learning. From differences in data sources, changes in the underlying processes generating the data, or new unseen environments the models have yet to encounter. All these issues can lead to performance degradation. From the success in using Cycle-consistent Generative Adversarial Networks(CycleGAN) to learn unpaired image-to-image mappings, we propose a new method to help alleviate the issues caused by domain shifts in images. The proposed model incorporates an adversarial loss to encourage realistic-looking images in the target domain, a cycle-consistency loss to learn an unpaired image-to-image mapping, and a semantic loss from a task network to improve the generator’s performance. The task network is con-currently trained with the generators on the generated images to improve downstream task performance on adapted images. By utilizing the power of CycleGAN, we can learn to classify images in the target domain without any target domain labels. In this research, we show that our model is successful on various unsupervised domain adaptation (UDA) datasets and can alleviate domain shifts for different adaptation tasks, like classification or semantic segmentation. In our experiments on standard classification, we were able to bring the models performance to near oracle level accuracy on a variety of different classification datasets. The semantic segmentation experiments showed that our model could improve the performance on the target domain, but there is still room for further improvements. We also further analyze where our model performs well and where improvements can be made.

4. Electronic Theses and Dissertations (ETDs) - Faculties submissions

Browse

Filters

Settings

Sort By

Results per page

Search Results