How do I transform a pandas groupby with a condition on the transformation?

Question

I have a dataframe containing session and bid data where there are three columns (of interest): user_id, event and date.

• user_id is just an ID to identify a user
• event is either bid or session
• date is a datetime object

Now what I want to do is add a column to my dataframe that is the date of the first bid. I have tried several ways of getting this to work but the issue is that it is of course very common that the user generated a session before they made a bid.

I have tried in several ways to get a filter to work but it does not seem to work like I think it should. From the documentation it says "Return a copy of a DataFrame excluding elements from groups that do not satisfy the boolean criterion specified by func." which sounds like what I want, ignore the events in the group that are session and not bid.

df['first bid date'] = df.groupby('user_id').filter(lambda x: x['event'] == 'bid')['date'].transform('min') 

When this did not work I tried to instead have the transform take a custom function, like this:

def custom_transform(group):
    return group[group['event'] == 'bid']['date'].min()


df['first bid date'] = df.groupby('user_id').['date'].transform(custom_transform)

But this does not work because the transform cannot access both the date and the event at the same time, seemingly no matter what I groupby.

Finally I tried to group by both the user_id and the event like this

df['first bid date'] = df.groupby(['user_id', 'event'])['date'].transform('min')

which kind of works but then I am left with having to change all of the first sessions to the first bid since there is now a first session and a first bid.

Any input to make this oneliner work? Seems like a combination of groupby, filter and transform should do the trick but I just can't crack it.

Thanks!

Answer 1

Idea is replace non matched values to missing values before transform , here by Series.where :

df['first bid date'] = (df.assign(date = df['date'].where(df['event'] == 'bid'))
                          .groupby('user_id')['date']
                          .transform('min'))

Answer 2

Here is some sample code with a dataframe to match the problem.

from io import StringIO

csv = StringIO("""index,uid,event,date
0,1,"bid",'2010-01-01'
1,1,"bid",'2013-01-01'
2,1,"session",'2009-01-01'
3,2,"session",'2010-01-01'
4,2,"bid",'2015-01-01'
5,2,"bid",'2017-01-01'""")

df = pd.read_csv(csv, index_col='index').reset_index(drop=True)

This alternate approach uses the merge function.

df.merge(df[df['event']=='bid'].groupby('uid')['date'].min(),
on='uid', suffixes=('','_first_bid'))

Which prints:

    uid  event    date        date_first_bid
0   1    bid      2010-01-01  2010-01-01
1   1    bid      2013-01-01  2010-01-01
2   1    session  2009-01-01  2010-01-01
3   2    session  2010-01-01  2015-01-01
4   2    bid      2015-01-01  2015-01-01
5   2    bid      2017-01-01  2015-01-01