253

I have a dataframe where some cells contain lists of multiple values. Rather than storing multiple values in a cell, I'd like to expand the dataframe so that each item in the list gets its own row (with the same values in all other columns). So if I have:

import pandas as pd
import numpy as np

df = pd.DataFrame(
    {'trial_num': [1, 2, 3, 1, 2, 3],
     'subject': [1, 1, 1, 2, 2, 2],
     'samples': [list(np.random.randn(3).round(2)) for i in range(6)]
    }
)

df
Out[10]: 
                 samples  subject  trial_num
0    [0.57, -0.83, 1.44]        1          1
1    [-0.01, 1.13, 0.36]        1          2
2   [1.18, -1.46, -0.94]        1          3
3  [-0.08, -4.22, -2.05]        2          1
4     [0.72, 0.79, 0.53]        2          2
5    [0.4, -0.32, -0.13]        2          3

How do I convert to long form, e.g.:

   subject  trial_num  sample  sample_num
0        1          1    0.57           0
1        1          1   -0.83           1
2        1          1    1.44           2
3        1          2   -0.01           0
4        1          2    1.13           1
5        1          2    0.36           2
6        1          3    1.18           0
# etc.

The index is not important, it's OK to set existing columns as the index and the final ordering isn't important.

cs95
  • 379,657
  • 97
  • 704
  • 746
Marius
  • 58,213
  • 16
  • 107
  • 105
  • 26
    From pandas 0.25 you can also use [`df.explode('samples')`](https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.explode.html#pandas.DataFrame.explode) to solve this. `explode` can only support exploding one column for now. – cs95 Aug 05 '19 at 20:50

10 Answers10

165

Pandas >= 0.25

Series and DataFrame methods define a .explode() method that explodes lists into separate rows. See the docs section on Exploding a list-like column.

df = pd.DataFrame({
    'var1': [['a', 'b', 'c'], ['d', 'e',], [], np.nan], 
    'var2': [1, 2, 3, 4]
})
df
        var1  var2
0  [a, b, c]     1
1     [d, e]     2
2         []     3
3        NaN     4

df.explode('var1')

  var1  var2
0    a     1
0    b     1
0    c     1
1    d     2
1    e     2
2  NaN     3  # empty list converted to NaN
3  NaN     4  # NaN entry preserved as-is

# to reset the index to be monotonically increasing...
df.explode('var1').reset_index(drop=True)

  var1  var2
0    a     1
1    b     1
2    c     1
3    d     2
4    e     2
5  NaN     3
6  NaN     4

Note that this also handles mixed columns of lists and scalars, as well as empty lists and NaNs appropriately (this is a drawback of repeat-based solutions).

However, you should note that explode only works on a single column (for now).

P.S.: if you are looking to explode a column of strings, you need to split on a separator first, then use explode. See this (very much) related answer by me.

cs95
  • 379,657
  • 97
  • 704
  • 746
150

A bit longer than I expected:

>>> df
                samples  subject  trial_num
0  [-0.07, -2.9, -2.44]        1          1
1   [-1.52, -0.35, 0.1]        1          2
2  [-0.17, 0.57, -0.65]        1          3
3  [-0.82, -1.06, 0.47]        2          1
4   [0.79, 1.35, -0.09]        2          2
5   [1.17, 1.14, -1.79]        2          3
>>>
>>> s = df.apply(lambda x: pd.Series(x['samples']),axis=1).stack().reset_index(level=1, drop=True)
>>> s.name = 'sample'
>>>
>>> df.drop('samples', axis=1).join(s)
   subject  trial_num  sample
0        1          1   -0.07
0        1          1   -2.90
0        1          1   -2.44
1        1          2   -1.52
1        1          2   -0.35
1        1          2    0.10
2        1          3   -0.17
2        1          3    0.57
2        1          3   -0.65
3        2          1   -0.82
3        2          1   -1.06
3        2          1    0.47
4        2          2    0.79
4        2          2    1.35
4        2          2   -0.09
5        2          3    1.17
5        2          3    1.14
5        2          3   -1.79

If you want sequential index, you can apply reset_index(drop=True) to the result.

update:

>>> res = df.set_index(['subject', 'trial_num'])['samples'].apply(pd.Series).stack()
>>> res = res.reset_index()
>>> res.columns = ['subject','trial_num','sample_num','sample']
>>> res
    subject  trial_num  sample_num  sample
0         1          1           0    1.89
1         1          1           1   -2.92
2         1          1           2    0.34
3         1          2           0    0.85
4         1          2           1    0.24
5         1          2           2    0.72
6         1          3           0   -0.96
7         1          3           1   -2.72
8         1          3           2   -0.11
9         2          1           0   -1.33
10        2          1           1    3.13
11        2          1           2   -0.65
12        2          2           0    0.10
13        2          2           1    0.65
14        2          2           2    0.15
15        2          3           0    0.64
16        2          3           1   -0.10
17        2          3           2   -0.76
Roman Pekar
  • 107,110
  • 28
  • 195
  • 197
  • Thanks, even the first step of applying to get each item in its own column is a huge help. I was able to come up with a slightly different way to do it, but there's still a fair few steps involved. Apparently this is not straightforward to do in Pandas! – Marius Dec 03 '14 at 12:30
  • 1
    Great answer. You can shorten it a bit by replacing `df.apply(lambda x: pd.Series(x['samples']),axis=1)` with `df.samples.apply(pd.Series)`. – Dennis Golomazov Jul 12 '17 at 01:33
  • 1
    Note to readers: This suffers horribly from performance issues. See [here](https://stackoverflow.com/a/48532692/4909087) for a much more performant solution using numpy. – cs95 Mar 13 '18 at 00:00
  • 2
    what's the solution when the number of samples is not the same for all rows? – SarahData Jun 08 '18 at 12:47
  • @SarahData Use `df.explode()` as shown [here.](https://stackoverflow.com/a/57122831/4909087) – cs95 Jul 20 '19 at 23:32
  • apply method is so slow... – hamidreza bina Aug 03 '22 at 05:51
78

UPDATE: the solution below was helpful for older Pandas versions, because the DataFrame.explode() wasn’t available. Starting from Pandas 0.25.0 you can simply use DataFrame.explode().

lst_col = 'samples'

r = pd.DataFrame({
      col:np.repeat(df[col].values, df[lst_col].str.len())
      for col in df.columns.drop(lst_col)}
    ).assign(**{lst_col:np.concatenate(df[lst_col].values)})[df.columns]

Result:

In [103]: r
Out[103]:
    samples  subject  trial_num
0      0.10        1          1
1     -0.20        1          1
2      0.05        1          1
3      0.25        1          2
4      1.32        1          2
5     -0.17        1          2
6      0.64        1          3
7     -0.22        1          3
8     -0.71        1          3
9     -0.03        2          1
10    -0.65        2          1
11     0.76        2          1
12     1.77        2          2
13     0.89        2          2
14     0.65        2          2
15    -0.98        2          3
16     0.65        2          3
17    -0.30        2          3

PS here you may find a bit more generic solution

UPDATE: some explanations: IMO the easiest way to understand this code is to try to execute it step-by-step:

in the following line we are repeating values in one column N times where N - is the length of the corresponding list:

In [10]: np.repeat(df['trial_num'].values, df[lst_col].str.len())
Out[10]: array([1, 1, 1, 2, 2, 2, 3, 3, 3, 1, 1, 1, 2, 2, 2, 3, 3, 3], dtype=int64)

this can be generalized for all columns, containing scalar values:

In [11]: pd.DataFrame({
    ...:           col:np.repeat(df[col].values, df[lst_col].str.len())
    ...:           for col in df.columns.drop(lst_col)}
    ...:         )
Out[11]:
    trial_num  subject
0           1        1
1           1        1
2           1        1
3           2        1
4           2        1
5           2        1
6           3        1
..        ...      ...
11          1        2
12          2        2
13          2        2
14          2        2
15          3        2
16          3        2
17          3        2

[18 rows x 2 columns]

using np.concatenate() we can flatten all values in the list column (samples) and get a 1D vector:

In [12]: np.concatenate(df[lst_col].values)
Out[12]: array([-1.04, -0.58, -1.32,  0.82, -0.59, -0.34,  0.25,  2.09,  0.12,  0.83, -0.88,  0.68,  0.55, -0.56,  0.65, -0.04,  0.36, -0.31])

putting all this together:

In [13]: pd.DataFrame({
    ...:           col:np.repeat(df[col].values, df[lst_col].str.len())
    ...:           for col in df.columns.drop(lst_col)}
    ...:         ).assign(**{lst_col:np.concatenate(df[lst_col].values)})
Out[13]:
    trial_num  subject  samples
0           1        1    -1.04
1           1        1    -0.58
2           1        1    -1.32
3           2        1     0.82
4           2        1    -0.59
5           2        1    -0.34
6           3        1     0.25
..        ...      ...      ...
11          1        2     0.68
12          2        2     0.55
13          2        2    -0.56
14          2        2     0.65
15          3        2    -0.04
16          3        2     0.36
17          3        2    -0.31

[18 rows x 3 columns]

using pd.DataFrame()[df.columns] will guarantee that we are selecting columns in the original order...

MaxU - stand with Ukraine
  • 205,989
  • 36
  • 386
  • 419
  • 3
    This should be the accepted answer. The currently accepted answer is much, much slower compared to this. – irene Aug 07 '18 at 05:29
  • 1
    I can't figure out how to fix this: TypeError: Cannot cast array data from dtype('float64') to dtype('int64') according to the rule 'safe' – Greg Oct 11 '18 at 16:54
  • 1
    This is the only answer that worked for me, out of the 10+ found during a full hour of searching the Stacks. Thanks MaxU – olisteadman Oct 26 '18 at 07:45
  • 1
    Note that this drops rows that have an empty list in `lst_col` entirely; to keep these rows and populate their `lst_col` with `np.nan`, you can just do `df[lst_col] = df[lst_col].apply(lambda x: x if len(x) > 0 else [np.nan])` before using this method. Evidently `.mask` won't return lists, hence the `.apply`. – Charles Davis Mar 14 '19 at 19:32
  • 1
    This is an excellent answer which should be the accepted one. Although, it's a black-magic level answer, and I, for one, would appreciate some explanation for what these steps in fact do. – ifly6 Apr 30 '19 at 17:19
  • Nice solution... but I would say `df[lst_col].apply(len)` instead of `df[lst_col].str.len()`. Just more clear since we're talking about lists not a strs. – travc May 08 '19 at 22:37
  • @travc, “apply” is too slow, so “.str.len()” is vectorized and therefore a more pandaic solution ;) – MaxU - stand with Ukraine May 08 '19 at 22:41
  • This is really hard to apply to non-random-generated data. – rjurney Jul 11 '19 at 23:00
  • THIS EXCLUDES rows when `lst_col` has empty list, people might also want to keep the rows with empty list. – BhishanPoudel Aug 31 '19 at 02:33
  • @MilkyWay001, you might want to check [this answer](https://stackoverflow.com/questions/12680754/split-explode-pandas-dataframe-string-entry-to-separate-rows/40449726?r=SearchResults&s=1%7C58.7107#40449726) ;) – MaxU - stand with Ukraine Aug 31 '19 at 07:36
12

you can also use pd.concat and pd.melt for this:

>>> objs = [df, pd.DataFrame(df['samples'].tolist())]
>>> pd.concat(objs, axis=1).drop('samples', axis=1)
   subject  trial_num     0     1     2
0        1          1 -0.49 -1.00  0.44
1        1          2 -0.28  1.48  2.01
2        1          3 -0.52 -1.84  0.02
3        2          1  1.23 -1.36 -1.06
4        2          2  0.54  0.18  0.51
5        2          3 -2.18 -0.13 -1.35
>>> pd.melt(_, var_name='sample_num', value_name='sample', 
...         value_vars=[0, 1, 2], id_vars=['subject', 'trial_num'])
    subject  trial_num sample_num  sample
0         1          1          0   -0.49
1         1          2          0   -0.28
2         1          3          0   -0.52
3         2          1          0    1.23
4         2          2          0    0.54
5         2          3          0   -2.18
6         1          1          1   -1.00
7         1          2          1    1.48
8         1          3          1   -1.84
9         2          1          1   -1.36
10        2          2          1    0.18
11        2          3          1   -0.13
12        1          1          2    0.44
13        1          2          2    2.01
14        1          3          2    0.02
15        2          1          2   -1.06
16        2          2          2    0.51
17        2          3          2   -1.35

last, if you need you can sort base on the first the first three columns.

behzad.nouri
  • 74,723
  • 18
  • 126
  • 124
  • 2
    This only works if you know a priori what the length of the lists will be and/or if they will all have the same length? – Chill2Macht Dec 06 '17 at 22:51
9

Trying to work through Roman Pekar's solution step-by-step to understand it better, I came up with my own solution, which uses melt to avoid some of the confusing stacking and index resetting. I can't say that it's obviously a clearer solution though:

items_as_cols = df.apply(lambda x: pd.Series(x['samples']), axis=1)
# Keep original df index as a column so it's retained after melt
items_as_cols['orig_index'] = items_as_cols.index

melted_items = pd.melt(items_as_cols, id_vars='orig_index', 
                       var_name='sample_num', value_name='sample')
melted_items.set_index('orig_index', inplace=True)

df.merge(melted_items, left_index=True, right_index=True)

Output (obviously we can drop the original samples column now):

                 samples  subject  trial_num sample_num  sample
0    [1.84, 1.05, -0.66]        1          1          0    1.84
0    [1.84, 1.05, -0.66]        1          1          1    1.05
0    [1.84, 1.05, -0.66]        1          1          2   -0.66
1    [-0.24, -0.9, 0.65]        1          2          0   -0.24
1    [-0.24, -0.9, 0.65]        1          2          1   -0.90
1    [-0.24, -0.9, 0.65]        1          2          2    0.65
2    [1.15, -0.87, -1.1]        1          3          0    1.15
2    [1.15, -0.87, -1.1]        1          3          1   -0.87
2    [1.15, -0.87, -1.1]        1          3          2   -1.10
3   [-0.8, -0.62, -0.68]        2          1          0   -0.80
3   [-0.8, -0.62, -0.68]        2          1          1   -0.62
3   [-0.8, -0.62, -0.68]        2          1          2   -0.68
4    [0.91, -0.47, 1.43]        2          2          0    0.91
4    [0.91, -0.47, 1.43]        2          2          1   -0.47
4    [0.91, -0.47, 1.43]        2          2          2    1.43
5  [-1.14, -0.24, -0.91]        2          3          0   -1.14
5  [-1.14, -0.24, -0.91]        2          3          1   -0.24
5  [-1.14, -0.24, -0.91]        2          3          2   -0.91
Marius
  • 58,213
  • 16
  • 107
  • 105
8

For those looking for a version of Roman Pekar's answer that avoids manual column naming:

column_to_explode = 'samples'
res = (df
       .set_index([x for x in df.columns if x != column_to_explode])[column_to_explode]
       .apply(pd.Series)
       .stack()
       .reset_index())
res = res.rename(columns={
          res.columns[-2]:'exploded_{}_index'.format(column_to_explode),
          res.columns[-1]: '{}_exploded'.format(column_to_explode)})
Charles Davis
  • 166
  • 2
  • 3
6

I found the easiest way was to:

  1. Convert the samples column into a DataFrame
  2. Joining with the original df
  3. Melting

Shown here:

    df.samples.apply(lambda x: pd.Series(x)).join(df).\
melt(['subject','trial_num'],[0,1,2],var_name='sample')

        subject  trial_num sample  value
    0         1          1      0  -0.24
    1         1          2      0   0.14
    2         1          3      0  -0.67
    3         2          1      0  -1.52
    4         2          2      0  -0.00
    5         2          3      0  -1.73
    6         1          1      1  -0.70
    7         1          2      1  -0.70
    8         1          3      1  -0.29
    9         2          1      1  -0.70
    10        2          2      1  -0.72
    11        2          3      1   1.30
    12        1          1      2  -0.55
    13        1          2      2   0.10
    14        1          3      2  -0.44
    15        2          1      2   0.13
    16        2          2      2  -1.44
    17        2          3      2   0.73

It's worth noting that this may have only worked because each trial has the same number of samples (3). Something more clever may be necessary for trials of different sample sizes.

Michael Silverstein
  • 1,653
  • 15
  • 17
4
import pandas as pd
df = pd.DataFrame([{'Product': 'Coke', 'Prices': [100,123,101,105,99,94,98]},{'Product': 'Pepsi', 'Prices': [101,104,104,101,99,99,99]}])
print(df)
df = df.assign(Prices=df.Prices.str.split(',')).explode('Prices')
print(df)

Try this in pandas >=0.25 version

Tapas
  • 75
  • 4
3

Very late answer but I want to add this:

A fast solution using vanilla Python that also takes care of the sample_num column in OP's example. On my own large dataset with over 10 million rows and a result with 28 million rows this only takes about 38 seconds. The accepted solution completely breaks down with that amount of data and leads to a memory error on my system that has 128GB of RAM.

df = df.reset_index(drop=True)
lstcol = df.lstcol.values
lstcollist = []
indexlist = []
countlist = []
for ii in range(len(lstcol)):
    lstcollist.extend(lstcol[ii])
    indexlist.extend([ii]*len(lstcol[ii]))
    countlist.extend([jj for jj in range(len(lstcol[ii]))])
df = pd.merge(df.drop("lstcol",axis=1),pd.DataFrame({"lstcol":lstcollist,"lstcol_num":countlist},
index=indexlist),left_index=True,right_index=True).reset_index(drop=True)
Khris
  • 3,132
  • 3
  • 34
  • 54
2

Also very late, but here is an answer from Karvy1 that worked well for me if you don't have pandas >=0.25 version: https://stackoverflow.com/a/52511166/10740287

For the example above you may write: 

data = [(row.subject, row.trial_num, sample) for row in df.itertuples() for sample in row.samples]
data = pd.DataFrame(data, columns=['subject', 'trial_num', 'samples'])

Speed test: 

%timeit data = pd.DataFrame([(row.subject, row.trial_num, sample) for row in df.itertuples() for sample in row.samples], columns=['subject', 'trial_num', 'samples'])

1.33 ms ± 74.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%timeit data = df.set_index(['subject', 'trial_num'])['samples'].apply(pd.Series).stack().reset_index()

4.9 ms ± 189 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit data = pd.DataFrame({col:np.repeat(df[col].values, df['samples'].str.len())for col in df.columns.drop('samples')}).assign(**{'samples':np.concatenate(df['samples'].values)})

1.38 ms ± 25 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

baitmbarek
  • 2,440
  • 4
  • 18
  • 26
DoRemy95
  • 614
  • 3
  • 19