Skip to content

Benchmarks

Here we will demonstrate some benchmarks against some alternatives.

Data

UCI Credit card dataset with 30k rows and 23 features.

import pandas as pd
from skorecard.datasets import load_credit_card
from sklearn.model_selection import train_test_split

from skorecard import Skorecard
from skorecard.pipeline.bucketing_process import BucketingProcess
from sklearn.pipeline import make_pipeline
from skorecard.bucketers.bucketers import DecisionTreeBucketer, OptimalBucketer

from time import time


data = load_credit_card(as_frame=True)
# data = pd.read_csv('UCI_Credit_Card.csv')
# cols = ["EDUCATION", "MARRIAGE", "LIMIT_BAL", "BILL_AMT1", "default"]
# data = data[cols]
# data.rename(columns={'default': 'y'}, inplace=True)


print(f"data shape: {data.shape}")

X_train, X_test, y_train, y_test = train_test_split(
    data.drop(["y"], axis=1), data[["y"]], test_size=0.25, random_state=42
)

data_train_opt, data_test_opt = train_test_split(data, test_size=0.25, random_state=42)
data_train_opt.head()


y_train = y_train.to_numpy().flatten()
y_test = y_test.to_numpy().flatten()
data shape: (30000, 24)

Experiment setup

from sklearn.metrics import roc_auc_score


def report_auc(clf, X_train, y_train, X_test, y_test):
    proba_train = clf.predict_proba(X_train)[:, 1]
    proba_test = clf.predict_proba(X_test)[:, 1]

    auc_train = round(roc_auc_score(y_train, proba_train), 4)
    auc_test = round(roc_auc_score(y_test, proba_test), 4)

    return auc_train, auc_test
from memo import memlist, time_taken

data = []


@memlist(data=data)
@time_taken()
def fit_eval_record(clf, name, opt=False):
    if opt:
        clf.fit(data_train_opt, data_train_opt["y"])
        proba_train = clf.predict_proba(data_train_opt)[:, 1]
        proba_test = clf.predict_proba(data_test_opt)[:, 1]

        auc_train = round(roc_auc_score(y_train, proba_train), 4)
        auc_test = round(roc_auc_score(y_test, proba_test), 4)

    else:
        clf.fit(X_train, y_train)
        auc_train, auc_test = report_auc(clf, X_train, y_train, X_test, y_test)

    return {"auc_train": auc_train, "auc_test": auc_test}

Skorecard is currently rather slow. A minor speed-up can be obtained by noting that both BucketingProcess and its pre-bucketers and bucketers compute identical bucket_tables and summaries: this is redundant when using a BucketingProcess. A boolean variable 'get_statistics' has been added to the bucketers to remove the calculation of these statistics. Below, a comparison is made to show the difference in speed this makes at the level of:

1) A single bucketer

2) A BucketingProcess

3) A full Scorecard pipeline

start_slow = time()
for i in range(10):
    bucketer_slow = DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=True)
    X_train_b1 = bucketer_slow.fit_transform(X_train, y_train)
end_slow = time()

print("Time for a single bucket when summary is computed:", end_slow - start_slow)

start = time()
for i in range(10):
    bucketer = DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=False)
    X_train_b2 = bucketer.fit_transform(X_train, y_train)
end = time()

print("Time for a single bucket when summary is not computed:", end - start)
Time for a single bucket when summary is computed: 39.013752937316895
Time for a single bucket when summary is not computed: 12.374780178070068

start_slow = time()
for i in range(5):
    clf_slow = BucketingProcess(
        prebucketing_pipeline=make_pipeline(
            DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=True)
        ),
        bucketing_pipeline=make_pipeline(OptimalBucketer(max_n_bins=10, min_bin_size=0.05, get_statistics=True)),
    )
    clf_slow.fit(X_train, y_train)
end_slow = time()
print("Time for a bucketing process when redundant summary is computed:", end_slow - start_slow)

start = time()
for i in range(5):
    clf = BucketingProcess(
        prebucketing_pipeline=make_pipeline(
            DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=False)
        ),
        bucketing_pipeline=make_pipeline(OptimalBucketer(max_n_bins=10, min_bin_size=0.05, get_statistics=False)),
    )
    clf.fit(X_train, y_train)
end = time()
print("Time for a bucketing process when redundant summary is not computed:", end - start)
Time for a bucketing process when redundant summary is computed: 66.75277090072632
Time for a bucketing process when redundant summary is not computed: 39.94823408126831

bucketing_process_slow = BucketingProcess(
    prebucketing_pipeline=make_pipeline(DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=True)),
    bucketing_pipeline=make_pipeline(OptimalBucketer(max_n_bins=10, min_bin_size=0.05, get_statistics=True)),
)
scorecard_slow = Skorecard(bucketing=bucketing_process_slow)

d_slow = fit_eval_record(scorecard_slow, name="skorecard.Scorecard")
print("Time for a scorecard model when redundant summary is computed:", d_slow["time_taken"])

bucketing_process = BucketingProcess(
    prebucketing_pipeline=make_pipeline(DecisionTreeBucketer(max_n_bins=100, min_bin_size=0.05, get_statistics=False)),
    bucketing_pipeline=make_pipeline(OptimalBucketer(max_n_bins=10, min_bin_size=0.05, get_statistics=False)),
)

scorecard = Skorecard(bucketing=bucketing_process)

d = fit_eval_record(scorecard, name="skorecard.Scorecard")
print("Time for a scorecard model when redundant summary is not computed:", d["time_taken"])
/Users/CK58LU/opt/anaconda3/envs/skorecard_env/lib/python3.9/site-packages/category_encoders/utils.py:21: FutureWarning: is_categorical is deprecated and will be removed in a future version. Use is_categorical_dtype instead.
  elif pd.api.types.is_categorical(cols):

Time for a scorecard model when redundant summary is computed: 19.25

/Users/CK58LU/opt/anaconda3/envs/skorecard_env/lib/python3.9/site-packages/category_encoders/utils.py:21: FutureWarning: is_categorical is deprecated and will be removed in a future version. Use is_categorical_dtype instead.
  elif pd.api.types.is_categorical(cols):

Time for a scorecard model when redundant summary is not computed: 13.61