target-likelihood-encoding
Basic idea: let's use target values to compute statistics for every sub-category of categorical features according to this formula:
smoothed likelihood = (fold_target_stat * nrows + global_val * alpha) / (nrows + alpha)
where:
- target_statistic - target statistic value across current fold,
- global_val - target statistic value across all train set,
- alpha - regularization value.
So if we have a rare subclass, it's target statistic will tend to the global value.
See the code for more info.
Usage
This kind of features leads to overfitting, so it must be created inside the cross-validation loop.
encoding = TLEncoding(n_splits=10, alpha=10, target='conversion')
encoding.fit(train)
tl_train = encoding.transform(train, mode='train')
tl_test = encoding.transform(test, mode='test')
Dependencies
- python 3.6
- numpy 1.12.1
- pandas 0.20.1