The purpose of this repository is to allow for a weights of evidence transformation of variables in a Pandas dataframe when performing a single class classification task. This type of transformation is often used in credit scoring models however can be used in a variety of applications. A weights of evidence transformation is typically performed on discrete or categorical variables, however this repository is able to handle numeric variables by providing an auto-grouping functionality.
When working with only categorical variables, the following can be done:
import WeightsOfEvidence
import pandas as pd
df = pd.read_csv('...')
target = '...'
columns = [col for col in df.columns if col != target]
woe = WeightsOfEvidence(df, columns, target)The transformed columns will then be available in woe.df and have a _woe suffix.
When numeric variables are available, grouping is performed. An example can be seen below:
woe = WeightsOfEvidence(df, columns, target, grouping_method='equal_bins')This will then create transformed columns as previously mentioned but also produce numeric variables with a _grouped suffix in woe.df.
Optimal splitting through decision tree classification can be performed using grouping_method='decision_tree'
The weights of evidence for all columns can be plotted using woe.plot_woe() to produce something like the following:
This example shows the plot for a categorical and numeric variable, where the numeric variable was grouped as previously described.
- No missing values should be present in either the columns to transform or the target variable. This class expects missings to be treated prior to calling.
- Future releases may incorporate additional grouping mechanisms.