Adult Census Income Dataset

Dataset where the prediction task is to determine whether a person makes over $50K a year. adult-census-income from kaggle.

NOTES:

TL;DR; XGBoost works best, and improves a little bit with target encoding

Logistic regression performs well on categorical variables only. XGBoost doesn't really improve here.
Without removing ordinal features, logistics regresion performs very badly because they need to be normalized first.
XGBoost performns very well with ordinal variables, as tree models don't need ordinal normalization.
Naive feature engineering combining categories in two improved xgb a little bit. Decreasing to 5 max depth improved a little bit extra.
Aplying target encoding over XGBoost improves the model a little bit further.

Train

cd src
conda activate ml
python create_folds.py
python -W ignore train.py [--model=xgb]  # [lr|rf|svd|xgb]
python -W ignore train_mean_target.py