Data Science, Diploma | Lighthouse Labs, Canada (November 2023) |
Applied Mathematics, PhD | Tshwane University of Technology, Pretoria, South Africa (November 2021) |
Mathematical Technology, MTech | Tshwane University of Technology, Pretoria, South Africa (July 2018) |
Mathematics, BSc | University of Lagos, Nigeria (October 2012) |
Transfer learning was used to retrain pretrained architectures to identify 17 classes of road signs. Custom dataset using one-shot learning was used as the problem is specific. The final layers of the EfficientNetB3, RestNet50, VGG16 architectures were tuned for training on the custom dataset and a preferred model was selected for deployment. The EfficientNetB3 model had an accuracy of 84% and was preferred because it is lightweight (for deployment purposes). Since there was no imbalance in the dataset, accuracy was the metric that was focused on.
A sample of the model in action is shown below
and you can access the deployed web app through this link. The entire repo is available here.
Twitter dataset was downloaded from a Twitter API (now X). The text file was in JSON format, positive and negative tweets had to be extracted using a for-loop. Preprocessing was done and data was converted using a vectorizer. The MultinomialNB, BernoulliNB, LogisticRegression, SVC, DecisionTreeClassifier, RandomForest models were tested and the preferred model was selected based on recall and ROC-AUC curve. The hyperparameters of the model were optimized and saved for deployment. The web app is accessible here. The project is available for viewing here.
Used Kmeans algorithm to identify crime hotspots using crime data for Los Angeles that was downloaded from the US Government open data website. Data cleaning was done in preparation for data analytics (in Tableau) and cluster analysis (in Python). Prescriptive policing model (crime segmentation) agrees with inferences from EDA. The project is available for viewing here.