Let's assume that I have the following data-frame:
df = pd.DataFrame({"id": [1, 1, 1, 2, 2], "nominal": [1, np.nan, 1, 1, np.nan], "numeric1": [3, np.nan, np.nan, 7, np.nan], "numeric2": [2, 3, np.nan, 2, np.nan], "numeric3": [np.nan, 2, np.nan, np.nan, 3], "date":[pd.Timestamp(2005, 6, 22), pd.Timestamp(2006, 2, 11), pd.Timestamp(2008, 9, 13), pd.Timestamp(2009, 5, 12), pd.Timestamp(2010, 5, 9)]})
As output, I want to get a data-frame, that will indicate the number of days that have passed since a non-nan value was seen for that column, for that id. If a column has a value for the corresponding date, or if a column doesn't have a value at the start for an new id, the value should be a 0. In addition, this is supposed to be computed only for the numeric columns. With that said, the output data-frame should be:
output_df = pd.DataFrame({"numeric1_delta": [0, 234, 1179, 0, 362], "numeric2_delta": [0, 0, 945, 0, 362], "numeric3_delta": [0, 0, 945, 0, 0]})
Looking forward to your answers!
You can groupby the cumsum of the non null and then subtract the first date:
In [11]: df.numeric1.notnull().cumsum()
Out[11]:
0 1
1 1
2 1
3 2
4 2
Name: numeric1, dtype: int64
In [12]: df.groupby(df.numeric1.notnull().cumsum()).date.transform(lambda x: x.iloc[0])
Out[12]:
0 2005-06-22
1 2005-06-22
2 2005-06-22
3 2009-05-12
4 2009-05-12
Name: date, dtype: datetime64[ns]
In [13]: df.date - df.groupby(df.numeric1.notnull().cumsum()).date.transform(lambda x: x.iloc[0])
Out[13]:
0 0 days
1 234 days
2 1179 days
3 0 days
4 362 days
Name: date, dtype: timedelta64[ns]
For multiple columns:
ncols = [col for col in df.columns if col.startswith("numeric")]
for c in ncols:
df[c + "_delta"] = df.date - df.groupby(df[c].notnull().cumsum()).date.transform('first')
Related
zed = pd.DataFrame(data = { 'date': ['2022-03-01', '2022-03-02', '2022-03-03', '2022-03-04', '2022-03-05'], 'a': [1, 5, 7, 3, 4], 'b': [3, 4, 9, 12, 5] })
How can the following dataframe be filtered to keep the earliest row (earliest == lowest date) for each of the 3 values 1, 5, 4 appearing in either column a or column b? In this example, the rows with dates '2022-03-01', '2022-03-02' would be kept as they are the lowest dates where each of the 3 values appears?
We have tried zed[zed.isin({'a': [1, 5, 4], 'b': [1, 5, 4]}).any(1)].sort_values(by=['date']) but this returns the incorrect result as it returns 3 rows.
Without reshape your dataframe, you can use:
idx = max([zed[['a', 'b']].eq(i).sum(axis=1).idxmax() for i in [1, 5, 4]])
out = zed.loc[:idx]
Output:
>>> out
date a b
0 2022-03-01 1 3
1 2022-03-02 5 4
You can reshape by DataFrame.stack, so possible filterin gby list with remove duplicates:
s = zed.set_index('date')[['a','b']].stack()
idx = s[s.isin([1, 5, 4])].drop_duplicates().index.remove_unused_levels().levels[0]
print (idx)
Index(['2022-03-01', '2022-03-02'], dtype='object', name='date')
out = zed[zed['date'].isin(idx)]
print (out)
date a b
0 2022-03-01 1 3
1 2022-03-02 5 4
Or filter first index value matching conditions, get unique values and select rows by DataFrame.loc:
L = [1, 5, 4]
idx = pd.unique([y for x in L for y in zed[zed[['a', 'b']].eq(x).any(axis=1)].index[:1]])
df = zed.loc[idx]
print (df)
date a b
0 2022-03-01 1 3
1 2022-03-02 5 4
I am trying to calculate a sum for each date field, however I only want to calculate the sum of IDs that are in both the current and next date, so a rolling comparison of IDs and then a groupby sum. Currently I have to loop over the dataframe which is very slow.
For example my df:
df = pd.DataFrame({
'Date': [1, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4],
'ID': [ 1, 2, 3, 4 , 2, 3, 4 , 2, 3, 4, 5, 1, 2, 3, 4],
'Value': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
})
Ideally I want to group the dataframe by Date and only sum the IDs that are common between two dates, for example below. However this is very slow.
tmpL = df.groupby('Date')['ID'].apply(list)
tmpV = df.groupby('Date')['Value'].sum()
for i in range(1, tmpL.shape[0]):
res = list(set(tmpL.iloc[i]) - set(tmpL.iloc[i - 1]))
v = df.loc[ df.ID.isin(res) & (df.Date == tmpL.index[i]), 'Value'].sum()
tmpV.iloc[i] = tmpV.iloc[i] - v
tmpV
Date
1 10
2 18
3 27
4 42
Name: Value, dtype: int64
Is there a way to do this in pandas without looping over the dataframe?
Use DataFrame.pivot_table with aggregate sum, compare for not equal with DataFrame.diff, and last passed to DataFrame.mask with sum:
df1 = df.pivot_table(index='Date', columns='ID', values='Value', aggfunc='sum')
s = df1.mask(df1.notna().diff().fillna(False)).sum(axis=1)
print (s)
Date
1 10.0
2 18.0
3 27.0
4 42.0
dtype: float64
First solution, I think slowier:
You can get all not matched sets by convert original to sets, then use Series.diff, Series.explode and get all matched values of original by DataFrame.merge, last aggregate sum and subtract:
tmpL = (df.groupby('Date')['ID'].apply(set)
.diff()
.explode()
.reset_index()
.merge(df)
.groupby('Date')['Value']
.sum())
tmpV = df.groupby('Date')['Value'].sum()
out = tmpV.sub(tmpL, fill_value=0)
print (out)
Date
1 10.0
2 18.0
3 27.0
4 42.0
Try:
df = df.pivot_table(index='Date', columns='ID', values='Value')#.reset_index()
condition = df.notna() & df.notna().shift(1)
condition.iloc[0,:]=True
print(df[condition].sum(axis=1))
Output:
Date
1 10.0
2 18.0
3 27.0
4 42.0
Suppose I have the following data frame
from pandas import DataFrame
Cars = { 'value': [10, 31, 661, 1, 51, 61, 551],
'action1': [1, 1, 1, 1, 1, 1, 1],
'price1': [ 12,0, 15,3, 0, 12,0],
'action2': [2, 2, 2, 2, 2, 2, 2],
'price2': [ 0, 16, 19, 0, 1, 10,0],
'action3': [3, 3, 3, 3, 3, 3, 3],
'price3': [ 14, 36, 9, 0, 0, 0,0]
}
df = DataFrame(Cars,columns= ['value', 'action1', 'price1', 'action2', 'price2', 'action3', 'price3'])
print (df)
How can I select randomly value (action and price) among 3 columns? As a result I want to have a dataframe that will look something like this one?
RandCars = {'value': [10, 31, 661, 1, 51, 61, 551],
'action': [1, 3, 1, 3, 1, 2, 2],
'price': [ 12, 36, 15, 0, 3, 10, 0]
}
df2 = DataFrame(RandCars, columns = ['value','action', 'price'])
print(df2)
Use:
#get columns names not starting by action or price
cols = df.columns[~df.columns.str.startswith(('action','price'))]
print (cols)
Index(['value'], dtype='object')
#convert filtered columns to 2 numpy arrays
arr1 = df.filter(regex='^action').values
arr2 = df.filter(regex='^price').values
#pandas 0.24+
#arr1 = df.filter(regex='^action').to_numpy()
#arr2 = df.filter(regex='^price').to_numpy()
i, c = arr1.shape
#create random positions of both DataFrames to new df
idx = np.random.choice(np.arange(c), i)
df3 = pd.DataFrame({'action': arr1[np.arange(len(df)), idx],
'price': arr2[np.arange(len(df)), idx]},
index=df.index)
print (df3)
action price
0 2 0
1 3 36
2 3 9
3 1 3
4 3 0
5 1 12
6 1 0
#add all another columns by join
df4 = df[cols].join(df3)
print (df4)
value action price
0 10 2 0
1 31 3 36
2 661 3 9
3 1 1 3
4 51 3 0
5 61 1 12
6 551 1 0
Create some data
cols = pd.MultiIndex.from_product([['what', 'why'], ['me', 'you']])
df = pd.DataFrame(columns=cols)
df.loc[0, :] = [1, 2, 3, 4]
What do we have?
In[8]: df
Out[8]:
what why
me you me you
0 1 2 3 4
Set one (or more) columns as index:
In[11]: df.set_index(('what', 'me'))
Out[11]:
what why
you me you
(what, me)
1 2 3 4
Let's reset that index:
In[12]: df.set_index(('what', 'me')).reset_index()
Out[12]:
(what, me) what why
you me you
0 1 2 3 4
And in particular,
In[13]: df.set_index(('what', 'me')).reset_index().columns
Out[13]:
MultiIndex(levels=[['what', 'why', ('what', 'me')], ['me', 'you', '']],
labels=[[2, 0, 1, 1], [2, 1, 0, 1]])
Is there any way to use these (multi) columns as indices without losing the column structure?
Suppose I use df.isnull().sum() and I get a count for all the 'NA' values in all the columns of df dataframe. I want to remove a column that has NA values above 'K'.
For eg.,
df = pd.DataFrame({'A': [1, 2.1, np.nan, 4.7, 5.6, 6.8],
'B': [0, np.nan, np.nan, 0, 0, 0],
'C': [0, 0, 0, 0, 0, 0.0],
'D': [5, 5, np.nan, np.nan, 5.6, 6.8],
'E': [0,np.nan,np.nan,np.nan,np.nan,np.nan],})
df.isnull().sum()
A 1
B 2
C 0
D 2
E 5
dtype: int64
Suppose I want to remove columns that have '2' and above number of NA values. How would be approach this problem? My output should be,
df.columns
A,C
Can anybody help me in doing this?
Thanks
Call dropna and pass axis=1 to drop column-wise and pass thresh=len(df)-K, what thresh does is it sets the minimum number of non-NaN values which is equal to the number of rows minus K NaN values
In [22]:
df.dropna(axis=1, thresh=len(df)-1)
Out[22]:
A C
0 1.0 0
1 2.1 0
2 NaN 0
3 4.7 0
4 5.6 0
5 6.8 0
If you just want the columns:
In [23]:
df.dropna(axis=1, thresh=len(df)-1).columns
Out[23]:
Index(['A', 'C'], dtype='object')
Or simply mask the counts output against the columns:
In [28]:
df.columns[df.isnull().sum() <2]
Out[28]:
Index(['A', 'C'], dtype='object')
Could do something like:
df = df.reindex(columns=[x for x in df.columns.values if df[x].isnull().sum() < threshold])
Which just builds a list of columns that match your requirement (fewer than threshold nulls), and then uses that list to reindex the dataframe. So if you set threshold to 1:
threshold = 1
df = pd.DataFrame({'A': [1, 2.1, np.nan, 4.7, 5.6, 6.8],
'B': [0, np.nan, np.nan, 0, 0, 0],
'C': [0, 0, 0, 0, 0, 0.0],
'D': [5, 5, np.nan, np.nan, 5.6, 6.8],
'E': ['NA', 'NA', 'NA', 'NA', 'NA', 'NA'],})
df = df.reindex(columns=[x for x in df.columns.values if df[x].isnull().sum() < threshold])
df.count()
Will yield:
C 6
E 6
dtype: int64
The dropna() function has a thresh argument that allows you to give the number of non-NaN values you require, so this would give you your desired output:
df.dropna(axis=1,thresh=5).count()
A 5
C 6
E 6
If you wanted just C & E, you'd have to change thresh to 6 in this case.