Pandas DataFrame: Rolling Set Union Aggregation over multiple Groups

Question

I have a DataFrame with a DateTimeIndex, a column I want to group by and a column containing sets of integers:

import pandas as pd

df = pd.DataFrame([['2018-01-01', 1, {1, 2, 3}],
                   ['2018-01-02', 1, {3}],
                   ['2018-01-03', 1, {3, 4, 5}],
                   ['2018-01-04', 1, {5, 6}],
                   ['2018-01-01', 2, {7}],
                   ['2018-01-02', 2, {8}],
                   ['2018-01-03', 2, {9}],
                   ['2018-01-04', 2, {10}]],
                  columns=['timestamp', 'group', 'ids'])

df['timestamp'] = pd.to_datetime(df['timestamp'])
df.set_index('timestamp', inplace=True)

            group        ids
timestamp                   
2018-01-01      1  {1, 2, 3}
2018-01-02      1        {3}
2018-01-03      1  {3, 4, 5}
2018-01-04      1     {5, 6}
2018-01-01      2        {7}
2018-01-02      2        {8}
2018-01-03      2        {9}
2018-01-04      2       {10}

Within each group I want to construct a rolling set union over the last x days. So assuming X=3 the result should be:

            group              ids
timestamp                   
2018-01-01      1        {1, 2, 3}
2018-01-02      1        {1, 2, 3}
2018-01-03      1  {1, 2, 3, 4, 5}
2018-01-04      1     {3, 4, 5, 6}
2018-01-01      2              {7}
2018-01-02      2           {7, 8}
2018-01-03      2        {7, 8, 9}
2018-01-04      2       {8, 9, 10}

From the answer to my previous question I got a good idea how to do this without the grouping, so I came up with this solution so far:

grouped = df.groupby('group')
new_df = pd.DataFrame()
for name, group in grouped:
    group['ids'] = [
        set.union(*group['ids'].to_frame().iloc(axis=1)[max(0, i-2): i+1,0])
        for i in range(len(group.index))
    ]
    new_df = new_df.append(group)

Which gives the correct result but looks quite clumsy and also gives the following warning:

SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: http://pandas.pydata.org/pandas-docs/stable/indexing.html#indexing-view-versus-copy

The documentation at the provided link does not really seem to fit my exact situation, though. (At least I can't make sense of it, in this context.)

My question: How can I improve this code to be clean, performant, and not throw the warning message?

Answer 1

As mentioned in the docs , don't use pd.DataFrame.append in a loop; doing so will be expensive.

Instead, use list and feed to pd.concat .

You can avoid SettingWithCopyWarning by creating copies of data within your list, ie avoid chained indexing via assign + iloc in a list comprehension:

L = [group.assign(ids=[set.union(*group.iloc[max(0, i-2): i+1, -1]) \
                       for i in range(len(group.index))]) \
     for _, group in df.groupby('group')]

res = pd.concat(L)

print(res)

            group              ids
timestamp                         
2018-01-01      1        {1, 2, 3}
2018-01-02      1        {1, 2, 3}
2018-01-03      1  {1, 2, 3, 4, 5}
2018-01-04      1     {3, 4, 5, 6}
2018-01-01      2              {7}
2018-01-02      2           {8, 7}
2018-01-03      2        {8, 9, 7}
2018-01-04      2       {8, 9, 10}