完全性

Pythonで学ぶデータクリーニング

Adel Nehme

VP of AI Curriculum, DataCamp

欠損データとは？

NA、nan、0、. などで表されます

技術的エラー

人的エラー

Airquality 例

import pandas as pd
airquality = pd.read_csv('airquality.csv')
print(airquality)

            Date  Temperature  CO2
987   20/04/2004         16.8  0.0
2119  07/06/2004         18.7  0.8
2451  20/06/2004        -40.0  NaN
1984  01/06/2004         19.6  1.8
8299  19/02/2005         11.2  1.2
...      ...             ...   ...

Airquality 例

import pandas as pd
airquality = pd.read_csv('airquality.csv')
print(airquality)

            Date  Temperature  CO2
987   20/04/2004         16.8  0.0
2119  07/06/2004         18.7  0.8
2451  20/06/2004        -40.0  NaN   <--
1984  01/06/2004         19.6  1.8
8299  19/02/2005         11.2  1.2
...      ...             ...   ...

Airquality 例

# 欠損値を取得
airquality.isna()

       Date  Temperature    CO2
987   False        False  False
2119  False        False  False
2451  False        False   True
1984  False        False  False
8299  False        False  False

Airquality 例

# 欠損の要約を取得
airquality.isna().sum()

Date             0
Temperature      0
CO2            366
dtype: int64

Missingno

欠損データの可視化と理解に便利なパッケージ

import missingno as msno
import matplotlib.pyplot as plt

# 欠損を可視化
msno.matrix(airquality)
plt.show()

Airquality 例

# 欠損と完全データを分離
missing = airquality[airquality['CO2'].isna()]
complete = airquality[~airquality['CO2'].isna()]

Airquality 例

# 完全データの記述統計
complete.describe()

       Temperature          CO2
count  8991.000000  8991.000000
mean     18.317829     1.739584
std       8.832116     1.537580
min      -1.900000     0.000000
...      ...        ...
max      44.600000    11.900000

# 欠損データの記述統計
missing.describe()

       Temperature  CO2
count   366.000000  0.0
mean    -39.655738  NaN
std       5.988716  NaN
min     -49.000000  NaN
...      ...        ...
max     -30.000000  NaN

Airquality 例

# 完全データの記述統計
complete.describe()

       Temperature          CO2
count  8991.000000  8991.000000
mean     18.317829     1.739584
std       8.832116     1.537580
min      -1.900000     0.000000
...      ...        ...
max      44.600000    11.900000

# 欠損データの記述統計
missing.describe()

       Temperature  CO2
count   366.000000  0.0
mean    -39.655738  NaN   <--
std       5.988716  NaN
min     -49.000000  NaN   <--
...      ...        ...
max     -30.000000  NaN   <--

sorted_airquality = airquality.sort_values(by = 'Temperature')
msno.matrix(sorted_airquality)
plt.show()

sorted_airquality = airquality.sort_values(by = 'Temperature')
msno.matrix(sorted_airquality)
plt.show()

欠損の種類

欠損への対処法

単純な方法：

欠損を削除
統計量で代入（平均・中央値・最頻値など）

高度な方法：

アルゴリズムで代入
機械学習モデルで代入

欠損への対処

airquality.head()

         Date  Temperature  CO2
0  05/03/2005          8.5  2.5
1  23/08/2004         21.8  0.0
2  18/02/2005          6.3  1.0
3  08/02/2005        -31.0  NaN
4  13/03/2005         19.9  0.1

欠損値の削除

# 欠損値を削除
airquality_dropped = airquality.dropna(subset = ['CO2'])
airquality_dropped.head()

         Date  Temperature  CO2
0  05/03/2005          8.5  2.5
1  23/08/2004         21.8  0.0
2  18/02/2005          6.3  1.0
4  13/03/2005         19.9  0.1
5  02/04/2005         17.0  0.8

統計量で置換

co2_mean = airquality['CO2'].mean()
airquality_imputed = airquality.fillna({'CO2': co2_mean})
airquality_imputed.head()

          Date  Temperature       CO2
0  05/03/2005          8.5  2.500000
1  23/08/2004         21.8  0.000000
2  18/02/2005          6.3  1.000000
3  08/02/2005        -31.0  1.739584
4  13/03/2005         19.9  0.100000

練習しましょう！

Pythonで学ぶデータクリーニング