Understanding Local Error in Regression Models (BART) by Lee Mabhena ’25

Understanding Local Error in Regression Models (BART) by Lee Mabhena ’25, Wednesday April 30, 1:00 – 1:50pm, North Science Building 015, Wachenheim, Statistics Thesis Defense

My thesis investigates how local factors—such as data sparsity, function gradient, and domain boundary location—affect the predictive performance of regression models. While global metrics like mean squared error (MSE) are standard for evaluation, they often obscure important local variations, especially in regions with steep gradients or limited data. Through controlled experiments using synthetic data, we compare four models: CART, Random Forests, BART, and Neural Networks. Results show that BART performs well out of the box, frequently outperforming RF and CART. It is more sensitive to changes in local regions, responding to gradient, domain position, and data sparsity, and it demonstrates stronger robustness to noise while effectively handling multi-dimensional data.