Features.py

import pandas as pd
import scipy.stats as stats

class CategoricalFeature():
    def __init__(self, df, feature):
        self.df = df
        self.feature = feature

    @property
    def df_lite(self):
        df_lite = self.df
        df_lite['bin'] = df_lite[self.feature].fillna('MISSING')
        return df_lite[['bin', 'label']]


class ContinuousFeature():
    def __init__(self, df, feature):
        self.df = df
        self.feature = feature
        self.bin_min_size = int(len(self.df) * 0.05)

    def __generate_bins(self, bins_num):
        df = self.df[[self.feature, 'label']]
        df['bin'] = pd.qcut(df[self.feature], bins_num, duplicates='drop') \
                    .apply(lambda x: x.left) \
                    .astype(float)
        return df

    def __generate_correct_bins(self, bins_max=20):
        for bins_num in range(bins_max, 1, -1):
            df = self.__generate_bins(bins_num)
            df_grouped = pd.DataFrame(df.groupby('bin') \
                                      .agg({self.feature: 'count',
                                            'label': 'sum'})) \
                                      .reset_index()
            r, p = stats.stats.spearmanr(df_grouped['bin'], df_grouped['label'])

            if (
                    abs(r)==1 and                                                        # check if woe for bins are monotonic
                    df_grouped[self.feature].min() > self.bin_min_size                   # check if bin size is greater than 5%
                    and not (df_grouped[self.feature] == df_grouped['label']).any()      # check if number of good and bad is not equal to 0
            ):
                break

        return df

    @property
    def df_lite(self):
        df_lite = self.__generate_correct_bins()
        df_lite['bin'].fillna('MISSING', inplace=True)
        return df_lite[['bin', 'label']]