最近參加了機器學習 百日馬拉松的活動，單純記錄下這100天python機器學習中每日覺得最有收穫的地方，如果有想參加這活動的朋友，真心推薦參加！此次機器學習-百日馬拉松的相關代碼放置於：https://github.com/hsuanchi/ML-100-days

相關文章：[百日馬拉松] 機器學習-資料清理https://www.maxlist.xyz/2019/03/03/ml_100days/

一. 標籤編碼 ( Label Encoding )

1. 類似於流⽔號，依序將新出現的類別依序編上新代碼，已出現的類別編上已使⽤的代碼
2. 確實能轉成分數，但缺點是分數的⼤⼩順序沒有意義

df_temp = pd.DataFrame()
for c in df.columns:
    df_temp[c] = LabelEncoder().fit_transform(df[c])

二. 獨熱編碼 ( One Hot Encoding )

1. 為了改良數字⼤⼩沒有意義的問題，將不同的類別分別獨立為⼀欄
2. 缺點是需要較⼤的記憶空間與計算時間，且類別數量越多時越嚴重

df_temp = pd.get_dummies(df)

三. 均值編碼 (Mean Encoding)

for col in object_features:
    print(col)
    df_mean = data.groupby([col]['Survived'].mean().reset_index()
    df_mean.columns = [col, f"{col}_mean"]
    data = pd.merge(data, df_mean, on=col, how='left')
    data = data.drop([col], axis=1)

四. 計數編碼 ( Counting )

1. 如果類別的⽬標均價與類別筆數呈正相關 ( 或負相關 )，也可以將筆數本⾝當成特徵

df_count = df.groupby('Cabin')['Name'].agg('size').reset_index()
df_count.columns = ['Cabin', 'Cabin_Count']
df = pd.merge(df, df_count, on='Cabin', how='left')
df.head()
df_temp = pd.DataFrame()
for col in object_features:
     df_temp[col] = LabelEncoder().fit_transform(df[col])
df_temp['Cabin_Hash'] = df['Cabin'].map(lambda x: hash(x) % 10)

五. 群聚編碼 (Group by Encoding)

1. 數值型特徵對⽂字型特徵最重要的特徵組合⽅式
2. 常⾒的有 mean, median, mode, max, min, count 等

mean_df = df.groupby(['Embarked'])['Fare'].mean().reset_index()
mode_df = df.groupby(['Embarked'])['Fare'].apply(lambda x: x.mode()[0]).reset_index()
median_df = df.groupby(['Embarked']['Fare'].median().reset_index()
max_df = df.groupby(['Embarked'])['Fare'].max().reset_index()
temp = pd.merge(mean_df, mode_df, how='left', on=['Embarked'])
temp = pd.merge(temp, median_df, how='left', on=['Embarked'])
temp = pd.merge(temp, max_df, how='left', on=['Embarked'])
temp.columns = ['Embarked', 'Fare_Mean', 'Fare_Mode', 'Fare_Median', 'Fare_Max']

六. 葉編碼(leaf encoding)

1. 將決策樹的葉點當做新特徵
2. 再用邏輯回歸合併預測

# 隨機森林擬合後
rf = RandomForestClassifier(n_estimators=20, min_samples_split=10, min_samples_leaf=5, 
                            max_features=4, max_depth=3, bootstrap=True)

將葉編碼 (*.apply) 結果做獨熱 / 邏輯斯迴歸
onehot = OneHotEncoder()
lr = LogisticRegression(solver='lbfgs', max_iter=1000)

rf.fit(train_X, train_Y)
onehot.fit(rf.apply(train_X))
lr.fit(onehot.transform(rf.apply(val_X)), val_Y)

七. 特徵雜湊 ( Feature Hash )

1. 特徵雜湊是⼀種折衷⽅案
2. 將類別由雜湊函數定應到⼀組數字
3. 調整雜湊函數對應值的數量
4. 在計算空間/時間與鑑別度間取折衷
5. 也提⾼了訊息密度, 減少無⽤的標籤

df_temp = pd.DataFrame()
for col in object_features:
    df_temp[col] = LabelEncoder().fit_transform(df[col])
df_temp['Cabin_Hash'] = df['Cabin'].map(lambda x: hash(x) % 10)

八. 空缺值：

df_m1 = df.fillna(-1)
df_0 = df.fillna(0)
df_md = df.fillna(df.median())

九. 標準化：

from sklearn.preprocessing import MinMaxScaler, StandardScaler
df_mm = MinMaxScaler().fit_transform(df)
df_sc = StandardScaler().fit_transform(df)

十. 離群值：

＃限制離群值在500-2000內
df['GrLivArea'] = df['GrLivArea'].clip(500, 2000)
sns.regplot(x = df['GrLivArea'], y=train_Y)
plt.show()

＃捨棄離群值
keep_indexs = (df['GrLivArea']> 500) & (df['GrLivArea']< 2000)
df = df[keep_indexs]
train_Y = train_Y[keep_indexs]
sns.regplot(x = df['GrLivArea'], y=train_Y)
plt.show()

十一. 去偏態：

＃原始
sns.distplot(df['Fare'])

# 對數去偏 (log1p)
df_fixed['Fare'] = np.log1p(df_fixed['Fare'])
sns.distplot(df_fixed['Fare'][:train_num])
plt.show()

# 分佈去偏(boxcox)
from scipy import stats
df_fixed['Fare'] = stats.boxcox(df_fixed['Fare']+1)[0]
sns.distplot(df_fixed['Fare'][:train_num])
plt.show()

# ⽅根去偏(sqrt)
df_fixed['Fare'] = stats.boxcox(df['Fare']+1, lmbda=0.5)
sns.distplot(df_fixed['Fare'][:train_num])
plt.show()

十二. 特徵工程流程：

1. 讀取資料
2. 分解重組與轉換
   • train+test=df
3. 特徵工程
   • Label Encoder
   • MinMax Encoder
4. 訓練模型與預測
5. 提交