Projects
Predicting Mortgage Default Using XGBoost for Classification
Date: December 11, 2024
This project explores predicting loan defaults using machine learning models like XGBoost and Decision Trees, with significant improvements in recall for defaulted loans.
This project tackled the challenges of predicting loan defaults using machine learning models. A baseline Decision Tree model was compared to XGBoost for advanced classification, and techniques like GridSearchCV and SMOTE were utilized for optimization and class balancing.
Highlights:
- Baseline Model (Decision Tree): Achieved 85% accuracy with limited recall for defaulted loans (7%).
- XGBoost Implementation: Improved accuracy to 87% and recall to 27%.
- Undersampling: Boosted recall to 66% for defaulted loans at the expense of reduced accuracy (75%).
Challenges and Contributions:
- Preprocessed a large, noisy dataset with missing and inconsistent data.
- Developed robust pipelines for hyperparameter tuning and cross-validation.
- Highlighted industry-wide issues in data collection practices, advocating for improvements.
Simulated Agent Populations with Diverse Reciprocity Pre-Dispositions
Date: January 12, 2025
This project models real-time helping behavior in a multi-agent environment, combining empirical data with simulated agent populations.
This project explores cooperative behavior in a multi-agent environment, combining empirical data analysis with computational modeling to understand decision-making dynamics.
Highlights:
- Data Source: Leveraged experimental data from a multi-agent environment (58,882 interactions).
- Agent Simulation: Modeled agent populations with varying reciprocity predispositions using logistic regression.
- Features: Key predictors included partner helping history and spatial proximity to resources.
- Results:
- High accuracy in predicting helping behavior.
- Diverse agent behaviors reflect real-world reciprocity dynamics.
- Explored the impact of cooperative and non-cooperative behaviors on game outcomes.
Contributions:
- Designed a novel agent generation process simulating individual variability in helping predispositions.
- Developed a logistic regression model for analyzing helping behavior predictors.
- Highlighted implications for cooperative strategies in dynamic environments.