Likelihood-Counts features

This algorithm is very helpful for handling the categorical features. It uses target values to compute smoothed likelihood (and counts) for every sub-category according to this formula:

smoothed likelihood = (mean(target) * nrows + global mean * alpha) / (nrows + alpha)

where:

• global mean - average target value across all train set,
• alpha - regularization value.

So if we have a rare subclass, it's likelihood will tend to the global mean value.

See the code for more info.

Installation

pip install lcfeatures

Usage

This kind of features leads to overfitting, so LC-Features must be created inside the cross-validation loop.

encoding = LCfeatures(TimeSeriesSplit(n_splits=5), modes=['mean', 'std', 'counter'], alpha=10, features='all', target='conversion'))
encoding.fit(train)
train = encoding.transform(train, mode='train')
test = encoding.transform(test, mode='test')

After that, new columns with suffix "_LC" will be created.

Dependencies

• python 3.6
• numpy 1.12.1
• pandas 0.20.1