はじめに

機械学習における不均衡二値分類において、undersampling + baggingがよい戦略だというのは様々なところで言われています。

今回は、この戦略をweightオプションの設定やoversamplingなど他の戦略と比較してみようと思います。

環境

Windows 10 Pro
Python 3.6.13
Jupyter Lab 6.4.3

評価手法・条件

LigthGBM
CV
しきい値: 0.5

dataset 1

credit card fraud
https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud

前準備

import pandas as pd
import numpy as np
from sklearn.model_selection import StratifiedKFold
import lightgbm as lgb

df = pd.read_csv('./input/creditcard.csv')
df.columns = [c.lower() for c in df.columns]

df['class'].value_counts()
# 0    284315
# 1       492

X = df.drop(['time', 'class'], axis=1).values
y = df['class'].values

前処理なし

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]
    
    # prepare datasets
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.679, 0.835, 0.749, 0.917

weightオプションの設定

lightGBMはweightオプションによりラベルの重みを変えることができます。

from sklearn.utils.class_weight import compute_sample_weight
train_weight = compute_sample_weight(class_weight='balanced', y=y_train).astype('float32')
lgb_train = lgb.Dataset(X_train, y_train, weight=train_weight)

以下、本コードです。

from sklearn.utils.class_weight import compute_sample_weight

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    train_weight = compute_sample_weight(class_weight='balanced', y=y_train).astype('float32')
    lgb_train = lgb.Dataset(X_train, y_train, weight=train_weight)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
1.0, 0.881, 0.825, 0.852, 0.913

undersampling

imblearnのRandomUndersamplerでundersamplingをしています。

from imblearn.under_sampling import RandomUnderSampler
rus = RandomUnderSampler(random_state=0, replacement=True)
X_train, y_train = rus.fit_resample(X_train, y_train)

以下、本コードです。

from imblearn.under_sampling import RandomUnderSampler

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    rus = RandomUnderSampler(random_state=0, replacement=True)
    X_train, y_train = rus.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.97, 0.05, 0.913, 0.094, 0.941

oversampling

imblearnのRandomOverSamplerでoversamplingをしています。

from imblearn.over_sampling import RandomOverSampler
ros = RandomOverSampler(random_state=0)
X_train, y_train = ros.fit_resample(X_train, y_train)

以下、本コードです。

from imblearn.over_sampling import RandomOverSampler

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    ros = RandomOverSampler(random_state=0)
    X_train, y_train = ros.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.879, 0.823, 0.85, 0.911

SMOTE

imblearnのSMOTEでoversamplingをしています。

from imblearn.over_sampling import SMOTE 
sm = SMOTE(random_state=0)
X_train, y_train = sm.fit_resample(X_train, y_train)

以下、本コードです。

from imblearn.over_sampling import SMOTE 

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    sm = SMOTE(random_state=0)
    X_train, y_train = sm.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.732, 0.854, 0.788, 0.927

ADASYN

imblearnのADASYNでoversamplingをします。

from imblearn.over_sampling import ADASYN 
ad = ADASYN(random_state=0)
X_train, y_train = ad.fit_resample(X_train, y_train)

以下、本コードです。

from imblearn.over_sampling import ADASYN 

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    ad = ADASYN(random_state=0)
    X_train, y_train = ad.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.73, 0.813, 0.769, 0.906

undersampling + bagging

いよいよ本題です。baggingで抽出したサンプルをundersamplingしてモデルを構築し、テストセットで予測し確率を出します。それを全てのbaggingの抜き取りに対してsoft votingし最終的な確率を出します。このプロセスをcvの各iterationごとに行います。

from imblearn.under_sampling import RandomUnderSampler

def undersample_bagg(X_train, X_eval, y_train, y_eval):
    # prepare datasets
    rus = RandomUnderSampler(replacement=True)
    X_train, y_train = rus.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)

    return y_pred_proba

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    n_bagging = 10
    preds = [undersample_bagg(X_train, X_eval, y_train, y_eval) for i in range(n_bagging)]
    y_pred_proba = sum(preds) / n_bagging

    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.974, 0.058, 0.911, 0.109, 0.942

undersampling + bagging + calibration

以下のブログでは、ダウンサンプリングによって予測時の確率にバイアスがかかる問題について論じています。

ダウンサンプリングによる予測確率のバイアス - sola

この不均衡データのダウンサンプリングによって、サンプル選択バイアスが生じることが Calibrating > Probability with Undersampling for Unbalanced Classification という論文で説明されています。
具体的には、少数派クラスの事前確率が大きくなります。一般的な問題設定では、正例のクラスが少数派クラスであるので、正例と予測される確率（事後確率）が大きくなります。
予測確率が重要な場合 *1 は特に、このバイアスの影響を除去しなければなりません。

そこで、本コードにおいてもバイアスの影響の除去を取り入れることにしました。変換の数式は以下です。

p = \frac{p_s}{p_s + \frac{(1 - p_s)}{\beta}}

$\beta$はダウンサンプリング率です。

from imblearn.under_sampling import RandomUnderSampler

def calibrate(prob, beta):
    return prob / (prob + (1 - prob) / beta)

def undersample_bagg(X_train, X_eval, y_train, y_eval):
    # prepare datasets
    rus = RandomUnderSampler(replacement=True)
    X_train, y_train = rus.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    # undersampling rate
    us_rate = sum(y_train == 1) / sum(y == 0)
    # calibrate probability
    y_pred_proba = calibrate(y_pred_proba, us_rate)

    return y_pred_proba

# stratified kfold split
kf = StratifiedKFold(n_splits=5, shuffle=True)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    n_bagging = 10
    preds = [undersample_bagg(X_train, X_eval, y_train, y_eval) for i in range(n_bagging)]
    y_pred_proba = sum(preds) / n_bagging

    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.852, 0.772, 0.81, 0.886

undersampling + calibration

undersamplingと確率補正の数式を取り入れたものです。

from imblearn.under_sampling import RandomUnderSampler

def calibrate(prob, beta):
    return prob / (prob + (1 - prob) / beta)

# stratified kfold split
kf = StratifiedKFold(n_splits=5)
oof = np.zeros(len(y))

# cv iterate through splits
for train_index, eval_index in kf.split(X, y):
    X_train, X_eval = X[train_index], X[eval_index]
    y_train, y_eval = y[train_index], y[eval_index]

    # prepare datasets
    rus = RandomUnderSampler(replacement=True)
    X_train, y_train = rus.fit_resample(X_train, y_train)
    lgb_train = lgb.Dataset(X_train, y_train)
    lgb_eval = lgb.Dataset(X_eval, y_eval, reference=lgb_train)

    # LightGBM hyperparameters
    lgbm_params = {
        'objective': 'binary',
        'metric': 'binary_logloss',
        'verbose': 0,
    }

    model = lgb.train(lgbm_params, lgb_train,
                        # validation data for the model
                        valid_sets=lgb_eval,
                        # train up to 10000 rounds
                        num_boost_round=10000,
                        # if the score doesn't increase for 10 rounds, stop training
                        early_stopping_rounds=10)

    # predict holdout with the trained model
    y_pred_proba = model.predict(X_eval, num_iteration=model.best_iteration)
    # undersampling rate
    us_rate = sum(y_train == 1) / sum(y == 0)
    # calibrate probability
    y_pred_proba = calibrate(y_pred_proba, us_rate)

    oof[eval_index] = (y_pred_proba > 0.5).astype(int)

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
print('accuracy_score, precision_score, recall_score, f1_score, roc_auc_score')
score_funcs = [accuracy_score, precision_score, recall_score, f1_score, roc_auc_score]
scores = [round(f(y, oof) ,3) for f in score_funcs]
print(', '.join(map(str, scores)))

accuracy_score, precision_score, recall_score, f1_score, roc_auc_score
0.999, 0.776, 0.695, 0.733, 0.847

まとめ

上記の実験についてスコアをまとめたものを作成しました。

condition	accuracy_score	precision_score	recall_score	f1_score	roc_auc_score
no treatment	0.999	0.716	0.823	0.766	0.911
set weight	0.999	0.86	0.825	0.842	0.912
undersampling	0.968	0.048	0.925	0.092	0.947
oversampling	0.999	0.836	0.821	0.829	0.91
SMOTE	0.999	0.632	0.85	0.725	0.924
ADASYN	0.999	0.679	0.827	0.746	0.913
undersampling + bagging	0.973	0.056	0.909	0.105	0.941
undersampling + bagging + calib	0.999	0.846	0.77	0.806	0.885
undersampling + calib	0.999	0.817	0.642	0.719	0.821

ここから分かることとしては以下です。

accuracyは全て>0.97と非常に高い（不均衡データなので当たり前）
SMOTE, ADASYNはoversamplingに比べてrecallは最大0.03程度上回るが、precisionは0.8台から0.6台へと大きく下がる
undersamplingおよびundersampling + baggingは、recallは>0.9と非常に高いがprecisionが0.05程度と大幅に低くなる
だが、両者は確率の補正を加えるとprecisionは0.8台、recallが0.6～0.8となる。しかしこれはweightのオプションを付けた状態に比べてprecision, recallともに劣る。

そのため、このデータセットに関しては、基本的にはweightのオプションを設定したもの、もしくはprecisionを完全に犠牲にしてrecallを高めたい場合はundersamplingもしくはそれにbaggingをつけたものがよいと考えます。

ですが、まだ迷いがあるため、もう一つデータセットをピックアップして検証してみたいと思います。

dataset 2

cancer dataset
https://www.kaggle.com/code/randyrose2017/for-beginners-using-keras-to-build-models/data

前処理

import pandas as pd
import numpy as np
from sklearn.model_selection import StratifiedKFold
import lightgbm as lgb

df = pd.read_csv('./input/cancer.csv')
df.columns = [c.lower() for c in df.columns]
df = df.replace('?', np.nan)

df = df.apply(pd.to_numeric)

df['biopsy'].value_counts()
# 0    803
# 1     55

# impute nan
for col, val in df.apply(lambda x: x.min(skipna=True)).to_dict().items():
    df[col][df[col].isnull()] = val - 1 # min - 1 for a GBDT model

検証のコード自体はdataset1と同じなので割愛します。

まとめ

上記の実験についてスコアをまとめたものを作成しました。

condition	accuracy_score	precision_score	recall_score	f1_score	roc_auc_score
no treatment	0.955	0.648	0.636	0.642	0.806
set weight	0.962	0.653	0.855	0.74	0.912
undersampling	0.955	0.6	0.873	0.711	0.916
oversampling	0.956	0.623	0.782	0.694	0.875
SMOTE	0.959	0.661	0.745	0.701	0.86
ADASYN	0.957	0.636	0.764	0.694	0.867
undersampling + bagging	0.962	0.649	0.873	0.744	0.92
undersampling + bagging + calib	0.95	0.688	0.4	0.506	0.694
undersampling + calib	0.952	0.706	0.436	0.539	0.712

これから分かることは以下です。

データセットによって、各不均衡データ向け処理の結果が変わることがある
accuracyは全て>0.95と非常に高い（不均衡データなので当たり前）
SMOTE, ADASYNはoversamplingに比べてrecallは最大0.04程度下回るが、precisionは0.04程度上がる。
undersamplingおよびundersampling + baggingは、recallは0.873と高く、precisionも0.6～0.65程度とである。
だが、両者は確率の補正を加えるとprecisionが最大0.1程度上昇するが、recallが0.4台に低下する

結論として、このデータセットの場合各手法に明確な優劣がつけづらいので、precisionとrecall、どちらが目的に沿うか検討して手法を選択するのがよいと考えます。

結論

手法の優劣がデータセットによって異なるため、一貫した結果が得られませんでした。と言っても2データセットしか試していませんが。undersampling + baggingに関しては、全てのデータセットで全てのスコアで優れた結果が出る、といった明確な結論が出ればよかったのですが、そうはなりませんでした。ただ一貫した傾向としては、recallに関して高めの値が出るがprecisionではそうでもないこと、確率の補正を行うとrecallは下がるがprecisionは上がることが挙げられると思います。理想としてはデータセットごとに全ての手法を試して目的に沿ったスコアが良い手法を選択することがありますが、かなりの労力がかかりますので、スコアが安定しているように見えるweightオプションを決め打ちで設定してしまうのも手かと思います。

コード

https://github.com/Dixhom/imbalanced-binary-labelled-data-comparison/tree/master

不均衡二値分類でundersampling + baggingは最適戦略か？

はじめに

環境

評価手法・条件

dataset 1

前準備

前処理なし

weightオプションの設定

undersampling

oversampling

SMOTE

ADASYN

undersampling + bagging

undersampling + bagging + calibration

undersampling + calibration

まとめ

dataset 2

前処理

まとめ

結論

コード