Weight of Evidence Demo

Simple implementation of 'weight of evidence binning' feature engineering method to improve perfomance of linear models in binary clasificaiton tasks.

Explanation

See admin/WoePresentation.pptx for more details on this methodology.

Prerequesites

Download Anaconda here

Setup

install & activate environment

conda env create -f environment.yml
conda activate weightOfEvidence

download test data 'German Credit'

cd weight_of_evidence
bash get_dataset.sh

run unit tests

cd weight_of_evidence
pytest .

Examples

fit sklearn pipeline on dataset

from sklearn.base import BaseEstimator, TransformerMixin
import pandas as pd
import numpy as np
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

import weight_of_evidence
DATA_PATH = "data/german_credit_data.csv"

COLUMN_NAMES = [
    "chk_acct",
    "duration",
    "credit_his",
    "purpose",
    "amount",
    "saving_acct",
    "present_emp",
    "installment_rate",
    "sex",
    "other_debtor",
    "present_resid",
    "property",
    "age",
    "other_install",
    "housing",
    "n_credits",
    "job",
    "n_people",
    "telephone",
    "foreign",
    "response",
]

data = pd.read_csv(DATA_PATH, sep=" ", names=COLUMN_NAMES)

data["response"] = data["response"] - 1

EXCLUDE_COLS = [
    "response",
]

X = data.drop(columns=EXCLUDE_COLS)
y = data["response"]

woebin_logit = Pipeline(
    steps=[
        ("tree_bin", weight_of_evidence.TreeBinner()),
        ("woe_scale", weight_of_evidence.LogitScaler()),
        ("standard_scale", StandardScaler()),
        ("log_reg_classifier", LogisticRegression(solver="lbfgs", max_iter=1e6)),
    ]
)

woebin_logit.fit(X, y)

create plots explaining fitted model

var_importance = weight_of_evidence.plot_feature_importance(
    X.columns, woebin_logit["log_reg_classifier"].coef_[0], n=5)


top_5 = var_importance.tail(5).var_names[::-1]

weight_of_evidence.plot_bins(
    X[top_5], y, woebin_logit["tree_bin"].splits_, space="log-odds"
)