2025ROC-AUC 0.8949 · precision 70.6% → 83.5%
U.S. Socioeconomic Income Risk Classifier
A census-scale income classifier benchmarking five algorithms with rigorous validation, feature dependency testing, and hyperparameter tuning.
- Python
- Scikit-learn
- Pandas
- Seaborn
- GridSearchCV
Preprocessed 41,000+ U.S. Census records with MinMax scaling, one-hot encoding, and custom binarization on skewed financial columns like capital gains and losses.
Method
- Chi-Square test for feature dependencies (p < 0.001)
- Five algorithms compared via stratified 5-fold cross-validation
- GridSearchCV tuning of the top model (C=100, â„“1 penalty)
Results
Final model reached ROC-AUC 0.8949, F1-score 0.6579, and 84.02% accuracy; precision improved from 70.6% to 83.5% after tuning while accuracy held steady.
Notable outcome
ROC-AUC 0.8949 · precision 70.6% → 83.5%