Skip to content

EBM benchmark with skorecard

This benchmark was adjusted from this notebook

# To run benchmark script, you will need to install XGBoost
# (pip install XGBoost)

import numpy as np
import pandas as pd
from sklearn.datasets import load_breast_cancer
import warnings

warnings.filterwarnings("ignore")


def load_breast_data():
    breast = load_breast_cancer()
    feature_names = list(breast.feature_names)
    X, y = pd.DataFrame(breast.data, columns=feature_names), breast.target
    dataset = {
        "problem": "classification",
        "full": {
            "X": X,
            "y": y,
        },
    }
    return dataset


def load_adult_data():
    df = pd.read_csv("https://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data", header=None)
    df.columns = [
        "Age",
        "WorkClass",
        "fnlwgt",
        "Education",
        "EducationNum",
        "MaritalStatus",
        "Occupation",
        "Relationship",
        "Race",
        "Gender",
        "CapitalGain",
        "CapitalLoss",
        "HoursPerWeek",
        "NativeCountry",
        "Income",
    ]
    train_cols = df.columns[0:-1]
    label = df.columns[-1]
    X_df = df[train_cols]
    y_df = df[label]

    dataset = {
        "problem": "classification",
        "full": {
            "X": X_df,
            "y": y_df,
        },
    }

    return dataset


def load_heart_data():
    # https://www.kaggle.com/ronitf/heart-disease-uci
    df = pd.read_csv(r"heart.csv")
    train_cols = df.columns[0:-1]
    label = df.columns[-1]
    X_df = df[train_cols]
    y_df = df[label]
    dataset = {
        "problem": "classification",
        "full": {
            "X": X_df,
            "y": y_df,
        },
    }

    return dataset


def load_credit_data():
    # https://www.kaggle.com/mlg-ulb/creditcardfraud
    df = pd.read_csv(r"creditcard.csv")
    train_cols = df.columns[0:-1]
    label = df.columns[-1]
    X_df = df[train_cols]
    y_df = df[label]
    dataset = {
        "problem": "classification",
        "full": {
            "X": X_df,
            "y": y_df,
        },
    }

    return dataset


def load_telco_churn_data():
    # https://www.kaggle.com/blastchar/telco-customer-churn
    df = pd.read_csv(r"WA_Fn-UseC_-Telco-Customer-Churn.csv")
    train_cols = df.columns[1:-1]  # First column is an ID
    label = df.columns[-1]
    X_df = df[train_cols]
    y_df = df[label]  # 'Yes, No'
    dataset = {
        "problem": "classification",
        "full": {
            "X": X_df,
            "y": y_df,
        },
    }

    return dataset
from sklearn.preprocessing import OneHotEncoder, FunctionTransformer, StandardScaler
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.model_selection import StratifiedShuffleSplit, cross_validate
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import OrdinalEncoder

from xgboost import XGBClassifier

from sklearn.linear_model import LogisticRegression

from interpret.glassbox import ExplainableBoostingClassifier

from skorecard import Skorecard

from optbinning import BinningProcess
from optbinning import Scorecard


def format_n(x):
    return f"{x:.3f}"


def process_model(clf, name, X, y, n_splits=3):
    # Evaluate model
    ss = StratifiedShuffleSplit(n_splits=n_splits, test_size=0.25, random_state=1337)
    scores = cross_validate(clf, X, y, scoring="roc_auc", cv=ss, n_jobs=-1, return_estimator=True)

    record = dict()
    record["model_name"] = name
    record["fit_time_mean"] = format_n(np.mean(scores["fit_time"]))
    record["fit_time_std"] = format_n(np.std(scores["fit_time"]))
    record["test_score_mean"] = format_n(np.mean(scores["test_score"]))
    record["test_score_std"] = format_n(np.std(scores["test_score"]))

    return record


def benchmark_models(dataset_name, X, y, ct=None, n_splits=3, random_state=1337):
    if ct is None:
        is_cat = np.array([dt.kind == "O" for dt in X.dtypes])
        cat_cols = X.columns.values[is_cat]
        num_cols = X.columns.values[~is_cat]

        cat_ohe_step = ("ohe", OneHotEncoder(sparse=False, handle_unknown="ignore"))

        cat_pipe = Pipeline([cat_ohe_step])
        num_pipe = Pipeline([("identity", FunctionTransformer())])
        transformers = [("cat", cat_pipe, cat_cols), ("num", num_pipe, num_cols)]
        ct = ColumnTransformer(transformers=transformers)

        cat_ord_step = ("ord_enc", OrdinalEncoder())
        cat_pipe = Pipeline([cat_ord_step])
        transformers = [("cat", cat_pipe, cat_cols), ("num", num_pipe, num_cols)]
        ot = ColumnTransformer(transformers=transformers)

    records = []

    summary_record = {}
    summary_record["dataset_name"] = dataset_name
    print()
    print("-" * 78)
    print(dataset_name)
    print("-" * 78)
    print(summary_record)
    print()

    pipe = Pipeline(
        [
            ("ct", ct),
            ("std", StandardScaler()),
            ("lr", LogisticRegression(random_state=random_state)),
        ]
    )
    record = process_model(pipe, "lr_ohe", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    pipe = Pipeline(
        [
            ("ot", ot),
            ("std", StandardScaler()),
            ("lr", LogisticRegression(max_iter=7000, random_state=random_state)),
        ]
    )
    record = process_model(pipe, "lr_ordinal", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    # Skorecard
    skorecard = Skorecard()
    record = process_model(skorecard, "skorecard", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    pipe = Pipeline(
        [
            ("ct", ct),
            # n_estimators updated from 10 to 100 due to sci-kit defaults changing in future versions
            ("rf-100", RandomForestClassifier(n_estimators=100, n_jobs=-1, random_state=random_state)),
        ]
    )
    record = process_model(pipe, "rf-100", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    pipe = Pipeline(
        [
            ("ct", ct),
            ("xgb", XGBClassifier(random_state=random_state, eval_metric="logloss")),
        ]
    )
    record = process_model(pipe, "xgb", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    # No pipeline needed due to EBM handling string datatypes
    ebm_inter = ExplainableBoostingClassifier(n_jobs=-1, random_state=random_state)
    record = process_model(ebm_inter, "ebm", X, y, n_splits=n_splits)
    print(record)
    record.update(summary_record)
    records.append(record)

    return records
results = []
n_splits = 3
from skorecard.datasets import load_uci_credit_card

X, y = load_uci_credit_card(return_X_y=True)
result = benchmark_models("UCI-creditcard", X, y, n_splits=n_splits)
results.append(result)

------------------------------------------------------------------------------
UCI-creditcard
------------------------------------------------------------------------------
{'dataset_name': 'UCI-creditcard'}

{'model_name': 'lr_ohe', 'fit_time_mean': '0.009', 'fit_time_std': '0.000', 'test_score_mean': '0.621', 'test_score_std': '0.023'}
{'model_name': 'lr_ordinal', 'fit_time_mean': '0.008', 'fit_time_std': '0.001', 'test_score_mean': '0.621', 'test_score_std': '0.023'}
{'model_name': 'skorecard', 'fit_time_mean': '1.489', 'fit_time_std': '0.044', 'test_score_mean': '0.627', 'test_score_std': '0.018'}
{'model_name': 'rf-100', 'fit_time_mean': '0.326', 'fit_time_std': '0.051', 'test_score_mean': '0.588', 'test_score_std': '0.013'}
{'model_name': 'xgb', 'fit_time_mean': '0.957', 'fit_time_std': '0.114', 'test_score_mean': '0.596', 'test_score_std': '0.005'}
{'model_name': 'ebm', 'fit_time_mean': '1.219', 'fit_time_std': '0.151', 'test_score_mean': '0.644', 'test_score_std': '0.012'}

dataset = load_breast_data()
result = benchmark_models("breast-cancer", dataset["full"]["X"], dataset["full"]["y"], n_splits=n_splits)
results.append(result)

------------------------------------------------------------------------------
breast-cancer
------------------------------------------------------------------------------
{'dataset_name': 'breast-cancer'}

{'model_name': 'lr_ohe', 'fit_time_mean': '0.014', 'fit_time_std': '0.001', 'test_score_mean': '0.994', 'test_score_std': '0.006'}
{'model_name': 'lr_ordinal', 'fit_time_mean': '0.015', 'fit_time_std': '0.001', 'test_score_mean': '0.994', 'test_score_std': '0.006'}
{'model_name': 'skorecard', 'fit_time_mean': '13.034', 'fit_time_std': '0.868', 'test_score_mean': '0.996', 'test_score_std': '0.004'}
{'model_name': 'rf-100', 'fit_time_mean': '0.253', 'fit_time_std': '0.002', 'test_score_mean': '0.992', 'test_score_std': '0.009'}
{'model_name': 'xgb', 'fit_time_mean': '0.196', 'fit_time_std': '0.014', 'test_score_mean': '0.992', 'test_score_std': '0.010'}
{'model_name': 'ebm', 'fit_time_mean': '5.866', 'fit_time_std': '1.070', 'test_score_mean': '0.995', 'test_score_std': '0.006'}

dataset = load_adult_data()
result = benchmark_models("adult", dataset["full"]["X"], dataset["full"]["y"], n_splits=n_splits)
results.append(result)
# 0.888

------------------------------------------------------------------------------
adult
------------------------------------------------------------------------------
{'dataset_name': 'adult'}

{'model_name': 'lr_ohe', 'fit_time_mean': '0.527', 'fit_time_std': '0.032', 'test_score_mean': '0.906', 'test_score_std': '0.003'}
{'model_name': 'lr_ordinal', 'fit_time_mean': '0.131', 'fit_time_std': '0.003', 'test_score_mean': '0.855', 'test_score_std': '0.002'}
{'model_name': 'skorecard', 'fit_time_mean': '8.047', 'fit_time_std': '0.025', 'test_score_mean': '0.888', 'test_score_std': '0.004'}
{'model_name': 'rf-100', 'fit_time_mean': '2.445', 'fit_time_std': '0.008', 'test_score_mean': '0.903', 'test_score_std': '0.002'}
{'model_name': 'xgb', 'fit_time_mean': '14.403', 'fit_time_std': '0.403', 'test_score_mean': '0.927', 'test_score_std': '0.001'}
{'model_name': 'ebm', 'fit_time_mean': '51.093', 'fit_time_std': '1.137', 'test_score_mean': '0.928', 'test_score_std': '0.002'}

dataset = load_telco_churn_data()
result = benchmark_models("telco_churn", dataset["full"]["X"], dataset["full"]["y"], n_splits=n_splits)
results.append(result)

------------------------------------------------------------------------------
telco_churn
------------------------------------------------------------------------------
{'dataset_name': 'telco_churn'}

{'model_name': 'lr_ohe', 'fit_time_mean': '9.091', 'fit_time_std': '0.030', 'test_score_mean': '0.809', 'test_score_std': '0.014'}
{'model_name': 'lr_ordinal', 'fit_time_mean': '0.064', 'fit_time_std': '0.002', 'test_score_mean': 'nan', 'test_score_std': 'nan'}
{'model_name': 'skorecard', 'fit_time_mean': '4.672', 'fit_time_std': '0.069', 'test_score_mean': '0.764', 'test_score_std': '0.014'}
{'model_name': 'rf-100', 'fit_time_mean': '7.118', 'fit_time_std': '0.016', 'test_score_mean': '0.824', 'test_score_std': '0.002'}
{'model_name': 'xgb', 'fit_time_mean': '104.738', 'fit_time_std': '0.969', 'test_score_mean': '0.825', 'test_score_std': '0.003'}
{'model_name': 'ebm', 'fit_time_mean': '36.052', 'fit_time_std': '3.779', 'test_score_mean': '0.852', 'test_score_std': '0.004'}

dataset = load_heart_data()
result = benchmark_models("heart", dataset["full"]["X"], dataset["full"]["y"], n_splits=n_splits)
results.append(result)

------------------------------------------------------------------------------
heart
------------------------------------------------------------------------------
{'dataset_name': 'heart'}

{'model_name': 'lr_ohe', 'fit_time_mean': '0.007', 'fit_time_std': '0.001', 'test_score_mean': '0.895', 'test_score_std': '0.030'}
{'model_name': 'lr_ordinal', 'fit_time_mean': '0.007', 'fit_time_std': '0.000', 'test_score_mean': '0.895', 'test_score_std': '0.030'}
{'model_name': 'skorecard', 'fit_time_mean': '2.148', 'fit_time_std': '0.013', 'test_score_mean': '0.911', 'test_score_std': '0.015'}
{'model_name': 'rf-100', 'fit_time_mean': '0.241', 'fit_time_std': '0.003', 'test_score_mean': '0.890', 'test_score_std': '0.008'}
{'model_name': 'xgb', 'fit_time_mean': '0.318', 'fit_time_std': '0.035', 'test_score_mean': '0.851', 'test_score_std': '0.018'}
{'model_name': 'ebm', 'fit_time_mean': '1.925', 'fit_time_std': '0.392', 'test_score_mean': '0.906', 'test_score_std': '0.011'}

records = [item for result in results for item in result]
record_df = pd.DataFrame.from_records(records)[["dataset_name", "model_name", "test_score_mean", "test_score_std"]]
record_df = record_df.sort_values(["dataset_name", "test_score_mean"], ascending=False)
print(
    record_df[record_df["model_name"].isin(["lr_ohe", "lr_ordinal", "rf-100", "skorecard", "xgb"])]
    .drop(["test_score_std"], axis=1)
    .to_markdown(tablefmt="github", showindex=False)
)
| dataset_name   | model_name   |   test_score_mean |
|----------------|--------------|-------------------|
| telco_churn    | lr_ordinal   |           nan     |
| telco_churn    | xgb          |             0.825 |
| telco_churn    | rf-100       |             0.824 |
| telco_churn    | lr_ohe       |             0.809 |
| telco_churn    | skorecard    |             0.764 |
| heart          | skorecard    |             0.911 |
| heart          | lr_ohe       |             0.895 |
| heart          | lr_ordinal   |             0.895 |
| heart          | rf-100       |             0.89  |
| heart          | xgb          |             0.851 |
| breast-cancer  | skorecard    |             0.996 |
| breast-cancer  | lr_ohe       |             0.994 |
| breast-cancer  | lr_ordinal   |             0.994 |
| breast-cancer  | rf-100       |             0.992 |
| breast-cancer  | xgb          |             0.992 |
| adult          | xgb          |             0.927 |
| adult          | lr_ohe       |             0.906 |
| adult          | rf-100       |             0.903 |
| adult          | skorecard    |             0.888 |
| adult          | lr_ordinal   |             0.855 |
| UCI-creditcard | skorecard    |             0.627 |
| UCI-creditcard | lr_ohe       |             0.621 |
| UCI-creditcard | lr_ordinal   |             0.621 |
| UCI-creditcard | xgb          |             0.596 |
| UCI-creditcard | rf-100       |             0.588 |