Feature Scaling with Partitions

[scidam]

Something like the following should work,

df['item_cnt_day'] = df.groupby('shop_id')['item_cnt_day'].transform(lambda x: (x-x.mean())/x.std())

However, if the group (a set of records with the same shop_id value) consist of one element only, this per-group scaling
will yield to NaN value (since x.std()= 0 if x is an array consisting of only one element).