Joshua Scantlebury
Joshua Scantlebury
Welcome to my portfolio! I enjoy various aspects of data science and here you’ll find a recap of my academic and personal projects.
📚 Projects
Enhancing Multiclass Classification with Random Forest and Bayesian Analysis
This project focuses on building and optimizing a multiclass classification model to address class imbalance and feature correlation. A Random Forest model, along with Logistic Regression, was used to train and predict the outcomes based on encoded and PCA-transformed data. Bayesian analysis was applied to evaluate the least important feature, providing insights into feature significance and model uncertainty.
Repository URL: https://github.com/joshdscan/Technical-Interview-Multi-Class-Classification
Key Highlights
- Implemented data preprocessing techniques such as outlier removal, scaling, and feature encoding to handle skewed distributions and missing data.
- The feature-selected Random Forest model achieved the highest weighted F1 score after hyperparameter tuning, outperforming other models with a 6.25% improvement.
- Bayesian analysis provided uncertainty estimation for feature importance, adding a probabilistic perspective to model performance.
Transforming Clinical Trial Labeling
This project focuses on addressing the high costs and low success rates associated with clinical trials. It develops a machine learning solution to automate the labeling of clinical trials based on termination reasons. By leveraging the transformer model, the project aims to reduce the manual curation burden and enhance the efficiency of identifying viable clinical trial candidates via multi-label classification.
Repository URL: https://github.com/joshdscan/CTWhyStopped/tree/main
Key Highlights
- Developed a deep learning solution for automating clinical trial labeling.
- Model displayed promise in cases where ample data was provided.
Automating Article Selection with Margin Insight
In this project, machine learning is used to automate the article selection process, focusing on articles likely to exceed 1400 shares for increased profitability. Utilizing a dataset of 40,000 articles with 61 features, including keyword counts and NLP metrics, this project employs a Random Forest model. This model boasts a 66% accuracy rate, surpassing the K-Nearest Neighbors model’s 58%.
Repository URL: https://github.com/joshdscan/Automating-Article-Selection
Key Highlights
- Automated article selection process targeting high-share articles for increased profitability.
- Predicted articles to collectively generate at least 25% more in average shares.
Exploring AirBnB Data for Maximizing Rental Value in New York City
This project was designed for a potential Airbnb host in The Big Apple. It utilizes a Tableau dashboard to guide strategic pricing decisions for maximizing rental property value. Addressing the unique market dynamics, the dashboard provides comparative insights with other properties, suggesting a balanced market price and evaluating renovation impacts.
Key Highlights
- Comparative market analysis suggesting a $190 nightly rate for maximized competitiveness.
- Evaluation of renovation impacts with a cautious approach to immediate property upgrades.
Reducing Credit Card Fraud at AJOS Bank
This project addresses the rise in fraud complaints within AJOS Bank’s credit card transactions. Four machine learning models were tested to analyze the bank’s data and identify the most effective method for detecting and preventing fraudulent transactions. The Random Forest model emerged as the most accurate, significantly reducing the potential financial losses and improving customer trust.
Repository URL: https://github.com/joshdscan/fraud-detection
Key Highlights
- Implemented oversampling to balance the dataset from a 99.6% no-fraud rate to a 60:40 ratio, significantly improving the model’s ability to detect fraud cases.
- The Random Forest model was selected as the best performer, with a 99.9% accuracy, 0.999 sensitivity (true positive rate), and 1 specificity (true negative rate), indicating its exceptional ability to identify fraudulent transactions without misclassifying legitimate ones.
Spare Parts Logistics Optimization Model
This project optimizes the logistics of spare parts distribution by minimizing total costs, including fixed costs for depots and transportation costs between suppliers, depots, and customers. Using Gurobi for mathematical modeling and Python for data handling, the model incorporates real life constraints such as supplier weight limits, depot capacity, and more to ensure efficient logistics operations while meeting service level standards.
Repository URL: https://github.com/joshdscan/Airplane-Spare-Parts-Logistics
Key Highlights
- Utilized Pandas and Gurobi to solve an optimization problem.
- Optimizing transportation costs can save over $150,000.
🔍 Interests
- Cinema: An avid movie watcher.
- Health and Wellness: Dedicated to advocating a balanced healthy lifestyle.
- Sports Analytics: Integrating passion for sports (particularly basketball) with analytics.
- Traveling: An explorer of diverse cultures and histories.
Feel free to reach out to me on LinkedIn or via email at jd.scantlebury@hotmail.com. Looking forward to connecting with you!