DataFrame のリンク

Pythonで学ぶデータクリーニング

Adel Nehme

VP of AI Curriculum, DataCamp

レコードリンケージ

DataFrame

census_A

             given_name  surname date_of_birth         suburb state  address_1
rec_id                                                                
rec-1070-org   michaela  neumann      19151111  winston hills   nsw  stanley street 
rec-1016-org   courtney  painter      19161214      richlands   vic  pinkerton circuit 
...

census_B

               given_name  surname date_of_birth             suburb state  address_1
rec_id                                                                      
rec-561-dup-0       elton      NaN      19651013         windermere   vic  light setreet 
rec-2642-dup-0   mitchell    maxon      19390212         north ryde   nsw  edkins street 
...

ここまでの作業

# recordlinkage をインポートし、全ペアを生成
import recordlinkage
indexer = recordlinkage.Index()
indexer.block('state')
full_pairs = indexer.index(census_A, census_B)


# 比較ステップ
compare_cl = recordlinkage.Compare()
compare_cl.exact('date_of_birth', 'date_of_birth', label='date_of_birth')
compare_cl.exact('state', 'state', label='state')
compare_cl.string('surname', 'surname', threshold=0.85, label='surname')
compare_cl.string('address_1', 'address_1', threshold=0.85, label='address_1')


potential_matches = compare_cl.compute(full_pairs, census_A, census_B)

現在の作業内容

候補一致

potential_matches

候補一致

potential_matches

候補一致

potential_matches

候補一致

potential_matches

高確度の一致

matches = potential_matches[potential_matches.sum(axis = 1) >= 3]
print(matches)

高確度の一致

matches = potential_matches[potential_matches.sum(axis = 1) >= 3]
print(matches)

インデックスを取得

matches.index

MultiIndex(levels=[['rec-1007-org', 'rec-1016-org', 'rec-1054-org', 'rec-1066-org', 
'rec-1070-org', 'rec-1075-org', 'rec-1080-org', 'rec-110-org', ...

# census_B のインデックスのみ取得
duplicate_rows = matches.index.get_level_values(1)
print(census_B_index)

Index(['rec-2404-dup-0', 'rec-4178-dup-0', 'rec-1054-dup-0', 'rec-4663-dup-0',
       'rec-485-dup-0', 'rec-2950-dup-0', 'rec-1234-dup-0', ... , 'rec-299-dup-0'])

DataFrame のリンク

# census_B の重複を検出
census_B_duplicates = census_B[census_B.index.isin(duplicate_rows)]


# census_B の新規行を抽出
census_B_new = census_B[~census_B.index.isin(duplicate_rows)]

# DataFrame をリンク
full_census = pd.concat([census_A, census_B_new])

# recordlinkage をインポートし、ペア生成と列比較
...
# 一致候補を生成
potential_matches = compare_cl.compute(full_pairs, census_A, census_B)


# 3 列以上が一致する行のみ抽出
matches = potential_matches[potential_matches.sum(axis = 1) >= 3]


# census_B の一致行インデックスのみ取得
duplicate_rows = matches.index.get_level_values(1)


# census_B の新規行を抽出
census_B_new = census_B[~census_B.index.isin(duplicate_rows)]


# DataFrame をリンク
full_census = pd.concat([census_A, census_B_new])

練習しましょう！

Pythonで学ぶデータクリーニング