ohe_target_as_feature.md

Effect

Up to 100% for model equals 1-ohe_target feature without modeling and generalizing anything.

Rules to avoid leakage

1. Use pipelines
2. Use arguments to skip one column during OHE
3. Check importances of features for trained models

Quick fix example

Instead of

X_train, y_train = (
    cc_apps_train_cat_encoding.iloc[:, :-1].values,
    cc_apps_train_cat_encoding.iloc[:, [-1]].values,
)

skip the column that was wrongly left during OHE

X_train = cc_apps_train_cat_encoding.iloc[:, :-2].values
y_train = cc_apps_train_cat_encoding.iloc[:, [-1]].values

Test

ohe_target_as_feature.py

Incorporation stage

Modeling

Locate in code

Fit, separation on target and features

Was met or loosely based on

datacamp "Predicting Credit Card Approvals" project