DataFrameの値をGroup毎に計算する時に，group keyが消えてしまうのを防ぐ

はじめに

こういうデータがあった時に

tmp = pd.DataFrame(dict(key=np.arange(10)//5, time=np.arange(10)%5, val1=np.random.randn(10),val2=np.random.randn(10)))
tmp

	key	time	val1	val2
0	0	0	-1.152069	0.788045
1	0	1	0.133470	0.347221
2	0	2	0.483643	0.487755
3	0	3	1.691687	1.426061
4	0	4	-0.575070	0.050923
5	1	0	-2.627809	-0.251222
6	1	1	0.668707	-1.490587
7	1	2	0.674961	0.623323
8	1	3	-1.788848	-0.915043
9	1	4	-1.027477	0.880744

以下のように'val1','val2'に対して
Group毎に処理をしようとするとGroupkeyが消えてしまう

tmp.groupby(['key'])[['val1','val2']].transform(lambda s: s)

	val1	val2
0	-1.152069	0.788045
1	0.133470	0.347221
2	0.483643	0.487755
3	1.691687	1.426061
4	-0.575070	0.050923
5	-2.627809	-0.251222
6	0.668707	-1.490587
7	0.674961	0.623323
8	-1.788848	-0.915043
9	-1.027477	0.880744

解決方法

消えてほしくないlabelはindexにしておく

tmp.set_index(['key','time']).groupby(['key'])[['val1','val2']].transform(lambda s: s)

		val1	val2
key	time
0	0	-1.152069	0.788045
	1	0.133470	0.347221
	2	0.483643	0.487755
	3	1.691687	1.426061
	4	-0.575070	0.050923
1	0	-2.627809	-0.251222
	1	0.668707	-1.490587
	2	0.674961	0.623323
	3	-1.788848	-0.915043
	4	-1.027477	0.880744

使用例

• key毎にnormalize

tmp.set_index(['key','time']).groupby(['key'])[['val1','val2']].transform(lambda s: (s-s.mean())/s.std())

		val1	val2
key	time
0	0	-1.169663	0.321366
	1	0.015803	-0.521662
	2	0.338717	-0.252906
	3	1.452722	1.541503
	4	-0.637580	-1.088302
1	0	-1.225673	-0.020611
	1	1.009441	-1.256779
	2	1.013681	0.851676
	3	-0.656838	-0.682719
	4	-0.140611	1.108433

• key毎にdiff

tmp.set_index(['key','time']).groupby(['key'])[['val1','val2']].diff()

		val1	val2
key	time
0	0	NaN	NaN
	1	1.285538	-0.440823
	2	0.350173	0.140534
	3	1.208045	0.938305
	4	-2.266757	-1.375138
1	0	NaN	NaN
	1	3.296516	-1.239366
	2	0.006253	2.113910
	3	-2.463809	-1.538365
	4	0.761371	1.795787

以上