Kaggleの臨床検査データセットを使ってみた⑤ ～不均衡データの取り扱い～

概要

Kaggleの血液検査データセットを使ってデータ分析をしてみた。
いろいろ試してみた結果、量が多くなったので分割します。
今回はその⑤ (全5回)

他の回はこちらから
①～モデルの性能比較をしてみた～
②～特徴選択をして、重要度を可視化してみた～
③～アンサンブル学習をしてみた～
④～外れ値の取り扱い～

使用したデータセット：Patient Treatment Classification (Electronic Health Record Dataset)
インドネシアの病院で集められた血液検査の結果から、患者に治療が必要かどうかを判定する

モデル

今回使用したモデルは以下の6種類

1. XGBoost
2. ランダムフォレスト
3. ロジスティック回帰
4. 決定木
5. k－近傍法

データ確認

データは血液検査の結果。
データの検査方法が分からなかったため、正常値は2016年度国立がん研究センターのデータをお借りしました。（※ 正常値は測定方法などにより若干のばらつきあり）

列No	検査値	和訳	正常値	高いと	低いと
1	HAEMATOCRIT	ヘマトクリット	男性：40.7～50.1 ％, 女性：35.1～44.4 ％	多血症など	貧血など
2	HAEMOGLOBINS	ヘモグロビン	男性：13.7～16.8 g/dl, 女性：11.6～14.8 g/dl	多血症など	貧血など
3	ERYTHROCYTE	赤血球	男性：4.35～5.55 x 106 /μＬ, 女性：3.86～4.92 x 106 /μＬ	多血症など	貧血など
4	LEUCOCYTE	白血球	3.3～8.6 x 103/μL	感染症・白血病など	一部感染症・膠原病・貧血など
5	THROMBOCYTE	血小板	158～348 x 103/μL	血小板血症・白血病・多血症など	貧血・紫斑病など
6	MCH	平均赤血球ヘモグロビン量	27.5～33.2 pg	巨赤芽球性貧血など	鉄欠乏性貧血など
7	MCHC	平均赤血球ヘモグロビン濃度	31.7～35.3 g/dL	脱水・多血症など	貧血など
8	MCV	平均赤血球容積	83.6～98.2 fL	巨赤芽球性貧血など	鉄欠乏性貧血など
9	AGE	年齢	―	―	―
10	SEX	性別	―	―	―
11	SOURCE	治療が必要か	out: 治療不要, in: 治療が必要	―	―

前処理は①と同じなので割愛

不均衡データの処理

ライブラリのインポート

from imblearn.over_sampling import SMOTE
from decimal import Decimal, ROUND_HALF_UP
import itertools

データ数確認

print(df_samp.shape)
print(df_samp["SOURCE"].value_counts())

>>
(4412, 35)
0    2628
1    1784
Name: SOURCE, dtype: int64

out: :2628データ、in: 1784データなので不均衡データとして処理する。
今回はSMOTEでOver-Samplingする。

SMOTEで水増し

水増し

data_y = df_samp["SOURCE"]
data_X = df_samp.drop("SOURCE",axis=1)
train_X, test_X, train_y, test_y = train_test_split(data_X, data_y, test_size=0.1)
sm = SMOTE()
x_resampled, y_resampled = sm.fit_resample(data_X, data_y.tolist())
# y_resampled.value_counts()
y_resampled = pd.DataFrame(y_resampled, columns=["SOURCE"])
y_resampled.value_counts()

>>
SOURCE
0         2628
1         2628
dtype: int64

水増ししたデータの性別が少数になったため、四捨五入して整数にする

性別を整数に

#0:男性 1:女性
int_column_list = ["SEX"]
for my_col in x_resampled.columns:
    # 整数にする
    if my_col in int_column_list:
        x_resampled[my_col] = x_resampled[my_col].map(lambda x: int(Decimal(str(x)).quantize(Decimal('0'), rounding=ROUND_HALF_UP)))

学習と結果

XGBoost

学習

# 変数を訓練用とテスト用に分割
train_X, test_X, train_y, test_y = train_test_split(x_resampled, y_resampled, test_size=0.2, random_state=42)
train_X, val_X, train_y, val_y = train_test_split(train_X, train_y, test_size=0.1, random_state=42)

# GBDTで学習
dtrain = xgb.DMatrix(train_X, label=train_y)
dval = xgb.DMatrix(val_X, label=val_y)
dtest = xgb.DMatrix(test_X)

params = {'objective': 'binary:logistic', 'silent':1, 'random_state': 71, 'eval_metric': 'auc'}
num_round = 25

watchlist = [(dtrain, 'train'), (dval, 'eval')]
GBDT_model_SM = xgb.train(params, dtrain, num_round, evals=watchlist)

val_pred = GBDT_model_SM.predict(dval)
val_y = val_y.values.tolist()
val_y = list(itertools.chain.from_iterable(val_y))
score_val = log_loss(val_y, val_pred)
print(f'logloss_val: {score_val: 4f}')

pred_y = GBDT_model_SM.predict(dtest)

結果

test_y = test_y.values.tolist()
pred_y = pred_y.tolist()
AUC_GBDT_SM = create_ROCcurve(test_y, pred_y)
tn_GBDT_SM, fp_GBDT_SM, fn_GBDT_SM, tp_GBDT_SM, specificity_GBDT_SM, recall_GBDT_SM, f1_GBDT_SM = create_cm(test_y, pred_y, True)
accuracy_GBDT_SM = accuracy_score(test_y, pred_y)
print(f'acc: {accuracy_GBDT_SM}')
AUC_list.append(["AUC_GBDT_SM:", np.round(AUC_GBDT_SM, 3)])
specificity_list.append(["specificity_GBDT_SM：", np.round(specificity_GBDT_SM, 3)])
recall_list.append(["recall_GBDT_SM：", np.round(recall_GBDT_SM, 3)])
f1_score_list.append(["f1_GBDT_SM：", np.round(f1_GBDT_SM, 3)])

結果

ランダムフォレスト

共通処理

train_X, test_X, train_y, test_y = train_test_split(x_resampled, y_resampled, test_size=0.2, random_state=42)
train_X, val_X, train_y, val_y = train_test_split(train_X, train_y, test_size=0.1, random_state=42)
train_X = np.asarray(train_X).astype(np.float32)
train_y = np.asarray(train_y).astype(np.float32)
val_X = np.asarray(val_X).astype(np.float32)
val_y = np.asarray(val_y).astype(np.float32)
test_X = np.asarray(test_X).astype(np.float32)
test_y = np.asarray(test_y).astype(np.float32)

学習と結果

RF_model_SM = RandomForestClassifier(max_depth=15, n_estimators=82)

RF_model_SM.fit(train_X, train_y)
RF_model_SM.score(test_X, test_y)

pred_y = RF_model_SM.predict(test_X)

AUC_RF_SM = create_ROCcurve(test_y, pred_y)
tn_RF_SM, fp_RF_SM, fn_RF_SM, tp_RF_SM, specificity_RF_SM, recall_RF_SM, f1_RF_SM = create_cm(test_y, pred_y)

AUC_list.append(["AUC_RF_SM", np.round(AUC_RF_SM, 3)])
specificity_list.append(["specificity_RF_SM", np.round(specificity_RF_SM, 3)])
recall_list.append(["recall_RF_SM", np.round(recall_RF_SM, 3)])
f1_score_list.append(["f1_RF_SM", np.round(f1_RF_SM, 3)])

結果

ロジスティック回帰

学習と結果

# ロジスティック回帰
LR_model_SM = LogisticRegression(max_iter=4500, C=79.42047819026727, random_state=40)

LR_model_SM.fit(train_X, train_y)
LR_model_SM.score(test_X, test_y)
pred_y = LR_model_SM.predict(test_X)

AUC_LR_SM = create_ROCcurve(test_y, pred_y)
tn_LR_SM, fp_LR_SM, fn_LR_SM, tp_LR_SM, specificity_LR_SM, recall_LR_SM, f1_LR_SM = create_cm(test_y, pred_y)

AUC_list.append(["AUC_LR_SM", np.round(AUC_LR_SM, 3)])
specificity_list.append(["specificity_LR_SM", np.round(specificity_LR_SM, 3)])
recall_list.append(["recall_LR_SM", np.round(recall_LR_SM, 3)])
f1_score_list.append(["f1_LR_SM", np.round(f1_LR_SM, 3)])

結果

決定木

学習と結果

DT_model_SM = DecisionTreeClassifier(max_depth=6)
DT_model_SM.fit(train_X, train_y)
DT_model_SM.score(test_X, test_y)

pred_y = DT_model_SM.predict(test_X)

AUC_DT_SM = create_ROCcurve(test_y, pred_y)
tn_DT_SM, fp_DT_SM, fn_DT_SM, tp_DT_SM, specificity_DT_SM, recall_DT_SM, f1_DT_SM = create_cm(test_y, pred_y)

AUC_list.append(["AUC_DT_SM", np.round(AUC_DT_SM, 3)])
specificity_list.append(["specificity_DT_SM", np.round(specificity_DT_SM, 3)])
recall_list.append(["recall_DT_SM", np.round(recall_DT_SM, 3)])
f1_score_list.append(["f1_DT_SM", np.round(f1_DT_SM, 3)])

結果

k-近傍法

学習と結果

KN_model_SM = KNeighborsClassifier(n_neighbors=15)
KN_model_SM.fit(train_X, train_y)
KN_model_SM.score(test_X, test_y)

pred_y = KN_model_SM.predict(test_X)

AUC_KN_SM = create_ROCcurve(test_y, pred_y)
tn_KN_SM, fp_KN_SM, fn_KN_SM, tp_KN_SM, specificity_KN_SM, recall_KN_SM, f1_KN_SM = create_cm(test_y, pred_y)

AUC_list.append(["AUC_KN_SM", np.round(AUC_KN_SM, 3)])
specificity_list.append(["specificity_KN_SM", np.round(specificity_KN_SM, 3)])
recall_list.append(["recall_KN_SM", np.round(recall_KN_SM, 3)])
f1_score_list.append(["f1_KN_SM", np.round(f1_KN_SM, 3)])

結果

スタッキング(ランダムフォレスト、ロジスティック回帰、決定木)

学習と結果

estimators = [('DT', DT_model_SM), ('RF', RF_model_SM), ('LR', LR_model_SM)]
stk_clf_SM = StackingClassifier(estimators=estimators, final_estimator=LogisticRegression(solver='liblinear'))
stk_clf_SM.fit(train_X, train_y)
stk_clf_SM.score(test_X, test_y)
pred_y = stk_clf_SM.predict(test_X)

# モデル評価
AUC_STK_SM = create_ROCcurve(test_y, pred_y)
tn_STK_SM, fp_STK_SM, fn_STK_SM, tp_STK_SM, specificity_STK_SM, recall_STK_SM, f1_STK_SM = create_cm(test_y, pred_y, True)
AUC_list.append(["AUC_Stacking_SM", np.round(AUC_STK_SM, 3)])
specificity_list.append(["specificity_Stacking_SM", np.round(specificity_STK_SM, 3)])
recall_list.append(["recall_Stacking_SM", np.round(recall_STK_SM, 3)])
f1_score_list.append(["f1_Stacking_SM", np.round(f1_STK_SM, 3)])

結果

スタッキング2(ランダムフォレスト、ロジスティック回帰、決定木、k-近傍法)

学習と結果

estimators = [('DT', DT_model_SM), ('RF', RF_model_SM), ('LR', LR_model_SM), ('KN', KN_model_SM)]
stk_clf_SM2 = StackingClassifier(estimators=estimators, final_estimator=LogisticRegression(solver='liblinear'))
stk_clf_SM2.fit(train_X, train_y)
stk_clf_SM2.score(test_X, test_y)
pred_y = stk_clf_SM2.predict(test_X)

# モデル評価
AUC_STK_SM2 = create_ROCcurve(test_y, pred_y)
tn_STK_SM2, fp_STK_SM2, fn_STK_SM2, tp_STK_SM2, specificity_STK_SM2, recall_STK_SM2, f1_STK_SM2 = create_cm(test_y, pred_y, True)
AUC_list.append(["AUC_Stacking_SM2", np.round(AUC_STK_SM2, 3)])
specificity_list.append(["specificity_Stacking_SM2", np.round(specificity_STK_SM2, 3)])
recall_list.append(["recall_Stacking_SM2", np.round(recall_STK_SM2, 3)])
f1_score_list.append(["f1_Stacking_SM2", np.round(f1_STK_SM2, 3)])

結果

外れ値除外 & 不均衡データ解消

共通の前処理

data_y_LOF = df_samp_LOF["SOURCE"]
data_X_LOF = df_samp_LOF.drop("SOURCE",axis=1)
train_X, test_X, train_y, test_y = train_test_split(data_X, data_y, test_size=0.1)
sm_LOF = SMOTE()
x_resampled_LOF, y_resampled_LOF = sm_LOF.fit_resample(data_X, data_y.tolist())
y_resampled_LOF = pd.DataFrame(y_resampled_LOF, columns=["SOURCE"])
y_resampled_LOF.value_counts()

int_column_list = ["SEX"]
for my_col in x_resampled_LOF.columns:
    # 整数にする
    if my_col in int_column_list:
        x_resampled_LOF[my_col] = x_resampled_LOF[my_col].map(lambda x: int(Decimal(str(x))
                                 .quantize(Decimal('0'), rounding=ROUND_HALF_UP)))

XGBoost

学習と結果

# 変数を訓練用とテスト用に分割
train_X, test_X, train_y, test_y = train_test_split(x_resampled_LOF, y_resampled_LOF, test_size=0.2, random_state=42)
train_X, val_X, train_y, val_y = train_test_split(train_X, train_y, test_size=0.1, random_state=42)

# GBDTで学習
dtrain = xgb.DMatrix(train_X, label=train_y)
dval = xgb.DMatrix(val_X, label=val_y)
dtest = xgb.DMatrix(test_X)

params = {'objective': 'binary:logistic', 'silent':1, 'random_state': 71, 'eval_metric': 'auc'}
num_round = 25

watchlist = [(dtrain, 'train'), (dval, 'eval')]
GBDT_model_SM_LOF = xgb.train(params, dtrain, num_round, evals=watchlist)

val_pred = GBDT_model_SM_LOF.predict(dval)
val_y = val_y.values.tolist()
val_y = list(itertools.chain.from_iterable(val_y))
score_val = log_loss(val_y, val_pred)
print(f'logloss_val: {score_val: 4f}')

pred_y = GBDT_model_SM_LOF.predict(dtest)

test_y = test_y.values.tolist()
pred_y = pred_y.tolist()
AUC_GBDT_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_GBDT_SM_LOF, fp_GBDT_SM_LOF, fn_GBDT_SM_LOF, tp_GBDT_SM_LOF, specificity_GBDT_SM_LOF, recall_GBDT_SM_LOF, f1_GBDT_SM_LOF = create_cm(test_y, pred_y, True)
accuracy_GBDT_SM_LOF = accuracy_score(test_y, pred_y)
print(f'acc: {accuracy_GBDT_SM_LOF}')
AUC_list.append(["AUC_GBDT_SM_LOF:", np.round(AUC_GBDT_SM_LOF, 3)])
specificity_list.append(["specificity_GBDT_SM_LOF：", np.round(specificity_GBDT_SM_LOF, 3)])
recall_list.append(["recall_GBDT_SM_LOF：", np.round(recall_GBDT_SM_LOF, 3)])
f1_score_list.append(["f1_GBDT_SM_LOF：", np.round(f1_GBDT_SM_LOF, 3)])

結果

ランダムフォレスト

共通処理

train_X, test_X, train_y, test_y = train_test_split(x_resampled, y_resampled, test_size=0.2, random_state=42)
train_X, val_X, train_y, val_y = train_test_split(train_X, train_y, test_size=0.1, random_state=42)
train_X = np.asarray(train_X).astype(np.float32)
train_y = np.asarray(train_y).astype(np.float32)
val_X = np.asarray(val_X).astype(np.float32)
val_y = np.asarray(val_y).astype(np.float32)
test_X = np.asarray(test_X).astype(np.float32)
test_y = np.asarray(test_y).astype(np.float32)

学習と結果

test_y = test_y.values.tolist()
pred_y = pred_y.tolist()
AUC_GBDT_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_GBDT_SM_LOF, fp_GBDT_SM_LOF, fn_GBDT_SM_LOF, tp_GBDT_SM_LOF, specificity_GBDT_SM_LOF, recall_GBDT_SM_LOF, f1_GBDT_SM_LOF = create_cm(test_y, pred_y, True)
accuracy_GBDT_SM_LOF = accuracy_score(test_y, pred_y)
print(f'acc: {accuracy_GBDT_SM_LOF}')
AUC_list.append(["AUC_GBDT_SM_LOF:", np.round(AUC_GBDT_SM_LOF, 3)])
specificity_list.append(["specificity_GBDT_SM_LOF：", np.round(specificity_GBDT_SM_LOF, 3)])
recall_list.append(["recall_GBDT_SM_LOF：", np.round(recall_GBDT_SM_LOF, 3)])
f1_score_list.append(["f1_GBDT_SM_LOF：", np.round(f1_GBDT_SM_LOF, 3)])

結果

ロジスティック回帰

学習と結果

LR_model_SM_LOF = LogisticRegression(max_iter=4500, C=79.42047819026727, random_state=40)

LR_model_SM_LOF.fit(train_X, train_y)
LR_model_SM_LOF.score(test_X, test_y)
pred_y = LR_model_SM_LOF.predict(test_X)

AUC_LR_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_LR_SM_LOF, fp_LR_SM_LOF, fn_LR_SM_LOF, tp_LR_SM_LOF, specificity_LR_SM_LOF, recall_LR_SM_LOF, f1_LR_SM_LOF = create_cm(test_y, pred_y)

AUC_list.append(["AUC_LR_SM_LOF", np.round(AUC_LR_SM_LOF, 3)])
specificity_list.append(["specificity_LR_SM_LOF", np.round(specificity_LR_SM_LOF, 3)])
recall_list.append(["recall_LR_SM_LOF", np.round(recall_LR_SM_LOF, 3)])
f1_score_list.append(["f1_LR_SM_LOF", np.round(f1_LR_SM_LOF, 3)])

結果

決定木

学習と結果

DT_model_SM_LOF = DecisionTreeClassifier(max_depth=6)
DT_model_SM_LOF.fit(train_X, train_y)
DT_model_SM_LOF.score(test_X, test_y)

pred_y = DT_model_SM_LOF.predict(test_X)

AUC_DT_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_DT_SM_LOF, fp_DT_SM_LOF, fn_DT_SM_LOF, tp_DT_SM_LOF, specificity_DT_SM_LOF, recall_DT_SM_LOF, f1_DT_SM_LOF = create_cm(test_y, pred_y)

AUC_list.append(["AUC_DT_SM_LOF", np.round(AUC_DT_SM_LOF, 3)])
specificity_list.append(["specificity_DT_SM_LOF", np.round(specificity_DT_SM_LOF, 3)])
recall_list.append(["recall_DT_SM_LOF", np.round(recall_DT_SM_LOF, 3)])
f1_score_list.append(["f1_DT_SM_LOF", np.round(f1_DT_SM_LOF, 3)])

結果

k-近傍法

学習と結果

KN_model_SM_LOF = KNeighborsClassifier(n_neighbors=15)
KN_model_SM_LOF.fit(train_X, train_y)
KN_model_SM_LOF.score(test_X, test_y)

pred_y = KN_model_SM_LOF.predict(test_X)

AUC_KN_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_KN_SM_LOF, fp_KN_SM_LOF, fn_KN_SM_LOF, tp_KN_SM_LOF, specificity_KN_SM_LOF, recall_KN_SM_LOF, f1_KN_SM_LOF = create_cm(test_y, pred_y)

AUC_list.append(["AUC_KN_SM_LOF", np.round(AUC_KN_SM_LOF, 3)])
specificity_list.append(["specificity_KN_SM_LOF", np.round(specificity_KN_SM_LOF, 3)])
recall_list.append(["recall_KN_SM_LOF", np.round(recall_KN_SM_LOF, 3)])
f1_score_list.append(["f1_KN_SM_LOF", np.round(f1_KN_SM_LOF, 3)])

結果

スタッキング2(ランダムフォレスト、ロジスティック回帰、決定木、k-近傍法)

学習と結果

estimators = [('DT', DT_model_SM_LOF), ('RF', RF_model_SM_LOF), ('LR', LR_model_SM_LOF), ('KN', KN_model_SM_LOF)]
stk_clf_SM_LOF = StackingClassifier(estimators=estimators, final_estimator=LogisticRegression(solver='liblinear'))
stk_clf_SM_LOF.fit(train_X, train_y)
stk_clf_SM_LOF.score(test_X, test_y)
pred_y = stk_clf_SM_LOF.predict(test_X)

results_SM_LOF = model_selection.cross_val_score(stk_clf, test_X, test_y, cv = kfold) 
scores[('4.Stacking', 'train_score')] = results_SM_LOF.mean()
scores[('4.Stacking', 'test_score')] = stk_clf_SM_LOF.score(test_X, test_y)

# モデル評価
print(pd.Series(scores).unstack())
AUC_STK_SM_LOF = create_ROCcurve(test_y, pred_y)
tn_STK_SM_LOF, fp_STK_SM_LOF, fn_STK_SM_LOF, tp_STK_SM_LOF, specificity_STK_SM_LOF, recall_STK_SM_LOF, f1_STK_SM_LOF = create_cm(test_y, pred_y, True)
AUC_list.append(["AUC_Stacking_SM_LOF", np.round(AUC_STK_SM_LOF, 3)])
specificity_list.append(["specificity_Stacking_SM_LOF", np.round(specificity_STK_SM_LOF, 3)])
recall_list.append(["recall_Stacking_SM_LOF", np.round(recall_STK_SM_LOF, 3)])
f1_score_list.append(["f1_Stacking_SM_LOF", np.round(f1_STK_SM_LOF, 3)])

結果

評価指標

評価指標

AUC_list = sorted(AUC_list, reverse=True, key=lambda x: x[1])
specificity_list = sorted(specificity_list, reverse=True, key=lambda x: x[1])
recall_list = sorted(recall_list, reverse=True, key=lambda x: x[1])
f1_score_list = sorted(f1_score_list, reverse=True, key=lambda x: x[1])

print("---ROC_AUC---")
for i in range(len(AUC_list)):
  print(AUC_list[i])
print("---特異度---")
for i in range(len(specificity_list)):
  print(specificity_list[i])
print("----感度----")
for i in range(len(recall_list)):
  print(recall_list[i])
print("----f1値----")
for i in range(len(f1_score_list)):
  print(f1_score_list[i])

ROC_AUC	特異度	感度	f1値
['AUC_GBDT_SM_LOF:', 0.861]	['specificity_RF', 0.859]	['recall_RF_SM_LOF', 0.78]	['f1_Stacking_SM_LOF', 0.793]
['AUC_GBDT_SM:', 0.859]	['specificity_GBDT：', 0.855]	['recall_Stacking_SM_LOF', 0.78]	['f1_GBDT_SM_LOF：', 0.792]
['AUC_GBDT:', 0.816]	['specificity_Stacking', 0.853]	['recall_GBDT_SM：', 0.78]	['f1_Stacking_SM2', 0.789]
['AUC_Stacking_SM_LOF', 0.795]	['specificity_LR', 0.837]	['recall_GBDT_SM_LOF：', 0.776]	['f1_Stacking_SM', 0.788]
['AUC_Stacking_SM', 0.791]	['specificity_LR_SM_LOF', 0.825]	['recall_Stacking_SM2', 0.774]	['f1_RF_SM_LOF', 0.787]
['AUC_Stacking_SM2', 0.79]	['specificity_LR_SM', 0.819]	['recall_Stacking_SM', 0.771]	['f1_GBDT_SM：', 0.787]
['AUC_RF_SM', 0.787]	['specificity_DT_SM', 0.815]	['recall_RF_SM', 0.769]	['f1_RF_SM', 0.785]
['AUC_RF_SM_LOF', 0.787]	['specificity_KN', 0.813]	['recall_KN_SM_LOF', 0.746]	['f1_LR_SM_LOF', 0.759]
['AUC_LR_SM_LOF', 0.771]	['specificity_Stacking_SM', 0.812]	['recall_DT_SM_LOF', 0.741]	['f1_DT_SM_LOF', 0.758]
['AUC_LR_SM', 0.766]	['specificity_GBDT_SM_LOF：', 0.812]	['recall_KN_SM', 0.737]	['f1_LR_SM', 0.754]
['AUC_DT_SM', 0.763]	['specificity_Stacking_SM_LOF', 0.81]	['recall_LR_SM_LOF', 0.716]	['f1_DT_SM', 0.751]
['AUC_DT_SM_LOF', 0.762]	['specificity_RF_SM', 0.806]	['recall_LR_SM', 0.712]	['f1_KN_SM_LOF', 0.735]
['AUC_RF', 0.755]	['specificity_Stacking_SM2', 0.806]	['recall_DT_SM', 0.711]	['f1_KN_SM', 0.734]
['AUC_Stacking', 0.755]	['specificity_RF_SM_LOF', 0.794]	['recall_Stacking', 0.658]	['f1_Stacking', 0.708]
['AUC_KN_SM', 0.73]	['specificity_GBDT_SM：', 0.794]	['recall_RF', 0.652]	['f1_RF', 0.707]
['AUC_KN_SM_LOF', 0.727]	['specificity_DT_SM_LOF', 0.783]	['recall_GBDT：', 0.634]	['f1_GBDT：', 0.692]
['AUC_LR', 0.723]	['specificity_DT', 0.737]	['recall_DT', 0.618]	['f1_LR', 0.666]
['AUC_KN', 0.702]	['specificity_KN_SM', 0.723]	['recall_LR', 0.61]	['f1_KN', 0.641]
['AUC_DT', 0.677]	['specificity_KN_SM_LOF', 0.708]	['recall_KN', 0.591]	['f1_DT', 0.625]

考察

不均衡データをオーバーサンプリングすることで、感度に大幅な改善が見られた。
またStacikngやGBDTなどのアンサンブル学習は単体での学習結果よりもよい結果が得られた。
加えて、外れ値を除外することで精度が改善されることがわかった。

参考資料

不均衡データの扱い方と評価指標！SmoteをPythonで実装して検証していく！
機械学習における不均衡データの扱いについて解説