[zerobase_데이터취업스쿨] 머신러닝_Credit Fraud Detection_(로그변환, IQR기준 이상치제거, 표준화스케일링)

6. amount 피처에 StandardScaler적용시켜보기(recall이 오히려 떨어졌다 precision은 약간 증가, f1은 감소, roc_auc도 감소)

scaler = StandardScaler()
amount_sc = scaler.fit_transform(np.array(X['Amount']).reshape(-1,1))

X_copy = X.iloc[:,:-1].copy()
X_copy['Amount_scaled'] = amount_sc

X_copy_train, X_copy_test, y_train, y_test = train_test_split(X_copy, y,random_state = 42, test_size = 0.3, stratify = y)

get_score_df(models, X_copy_train, y_train, X_copy_test, y_test)

C:\Users\kd010\miniconda3\Lib\site-packages\sklearn\linear_model\_logistic.py:1222: UserWarning: 'n_jobs' > 1 does not have any effect when 'solver' is set to 'liblinear'. Got 'n_jobs' = 16.
  warnings.warn(

[LightGBM] [Info] Number of positive: 344, number of negative: 199020
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.009484 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 7395
[LightGBM] [Info] Number of data points in the train set: 199364, number of used features: 29
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

7. amount 피처 로그 변환 해보기 lop1p() = log(1+x) 활용

• LogisticRegression의 recall이 약간 증가한것 0.01정도 빼고는 별 변화 없음
• 로그 변환 결과 분포가 훨씬 균일해짐

X_log = X.iloc[:,:-1].copy()
amount_log = np.log1p(np.array(X['Amount']).reshape(-1,1))

X_log['Amount_log'] = amount_log

X_log_train, X_log_test, y_train, y_test = train_test_split(X_log, y, stratify = y, random_state = 42, test_size = 0.2)

get_score_df(models, X_log_train, y_train, X_log_test, y_test)

C:\Users\kd010\miniconda3\Lib\site-packages\sklearn\linear_model\_logistic.py:1222: UserWarning: 'n_jobs' > 1 does not have any effect when 'solver' is set to 'liblinear'. Got 'n_jobs' = 16.
  warnings.warn(

[LightGBM] [Info] Number of positive: 394, number of negative: 227451
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.009853 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 7395
[LightGBM] [Info] Number of data points in the train set: 227845, number of used features: 29
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

sns.histplot(X_log, x='Amount_log')

<Axes: xlabel='Amount_log', ylabel='Count'>

8. 대략 boxplot으로 이상치 파악하기

fig, ax = plt.subplots(2,1,figsize=(12,7))
sns.boxplot(X.iloc[:,:10],ax = ax[0])
sns.boxplot(X.iloc[:,10:-1],ax=ax[1])
plt.show()

9. V20, V21, V23의 이상치가 많아보이므로 이상치 없애보기

def set_outliers(data, column):

    fraud = data[data['Class'] == 0][column]
    q1 = np.quantile(fraud, 0.25)
    q3 = np.quantile(fraud, 0.75)

    iqr = q3 - q1
    lower_bound = q1 - 1.5*iqr
    upper_bound = q3 + 1.5*iqr
    cond1 = fraud < lower_bound
    cond2 = fraud > upper_bound
    indexes = fraud[cond1 | cond2].index
    
    return indexes.tolist()

X_cols = X.columns.tolist()

outliers_indexes = []
for col in ['V20','V21','V23']:
    outliers_indexes.extend(set_outliers(raw_df, col))

outliers = list(set([i for i in outliers_indexes[::1]]))

print(len(outliers))

38892

new_df = raw_df.drop(index=outliers)

new_df

10. Class가 0인 것(Not Fraud) 중 V20, V21, V23칼럼의 이상치 제거만 하고 성능평가하기

• 나의 논리는 Not Fraud로 라벨링 된 것 중 이상치들은 사실 Fraud로 분류되어야 하지 않았을까 라는 생각에서 V20, V21, V23의 oulier를 iqr에 기반하여 제거하였다
• 제거한 후 성능을 보면 현재까지 시도(raw데이터, amount칼럼 standardscaling, amount칼럼 log변환)중 가장 높은 recall과 precision, f1score 모두 압도적으로 높다

X = new_df.drop(['Time','Class'],axis=1)
y = new_df['Class']

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2, stratify = y, random_state = 42)

get_score_df(models, X_train, y_train, X_test, y_test)

[LightGBM] [Info] Number of positive: 394, number of negative: 196338
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.007124 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 7395
[LightGBM] [Info] Number of data points in the train set: 196732, number of used features: 29
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

11. Outlier 제거된 데이터에서 Amount열을 log변환 한 성능 평가

• V20, V21, V23 열 중에 Class가 0인것 중 iqr을 기준으로 outlier를 제거한 것과 그 데이터를 기반으로 Amount칼럼을 log변환한 성능이 동일함

X_log = X.iloc[:,:-1].copy()
amount_log = np.log1p(X['Amount'])
X_log['Amount_log'] = amount_log

X_log_train, X_log_test, y_train, y_test = train_test_split(X_log, y, test_size=0.2, random_state=42, stratify=y)

get_score_df(models, X_log_train, y_train, X_log_test, y_test)

[LightGBM] [Info] Number of positive: 394, number of negative: 196338
[LightGBM] [Info] Auto-choosing col-wise multi-threading, the overhead of testing was 0.007568 seconds.
You can set `force_col_wise=true` to remove the overhead.
[LightGBM] [Info] Total Bins 7395
[LightGBM] [Info] Number of data points in the train set: 196732, number of used features: 29
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf