College Outcomes Prediction

Regression analysis predicting college graduate earnings using anonymized data from the U.S. Department of Education College Scorecard.

Competition

DrivenData Machine Learning Competition evaluating 17,107 post-collegiate outcomes across 297 variables to construct predictions for 9,912 test samples.

Key Findings

Notable discoveries included that “public, nonprofit” school status and Ph.D. program presence were strong positive income indicators, while culinary programs and “private, for-profit” status correlated with lower earnings.

Approach

Applied generalized stacking of seven tree-based models (AdaBoost, Extra Trees, Random Forest, Gradient Boosted Decision Trees, LightGBM, XGBoost) with quadratic feature interactions, missing data interpolation, outlier management, and ordinal ranking of states and demographics.