Before training ML models from multiple datasets, we need to understand how to relate datasets together. We’ll cover how to look at relationships in data using
Before we begin
In this guide we’ll go over one of the most core concepts in AI that most search for today, how to combine dataframes with
. Welcome to the 2nd part of the “Product developers’ guide to customizing data for AI”. In this article we’ll go over intermediate concepts and run through examples using
in Pandas. We’ll look at the relationships between the email content dataset and user email dataset. By the end of this series, you’ll be ready to tackle designing datasets in order to begin training machine learning models.
Before we begin
In this guide, we’ll reuse ouremail_content
dataset and include theuser_email
dataset. Like all other guides, we’ll be using Google Collab and start by importing those datasets into dataframes. There will be references back topart 1
where we looked at folding, and basic set relations and continue off from there. If you aren’t familiar with transforming or sorting data, please read our introductoryguide
Read from the start
Previously, we covered unions and intersections by concatenating dataframes. Building off of it, we introduce relative joins to always include a certain set.
Another type of join is based on a certain side of the dataset. In set theory this is called a relative join, and contains everything on 1 side. In our case of having 2 datasets, we have 2 sides, the left side and the right side.
Depending on the side, it must include all values from the set (Source: CodeProject)
In Pandas, the
function can be used together with keys, or unique pairs used to form the relationships seen in above. These are demonstrated best when there is data that is unique to both dataframes, in other words when the data doesn’t always exist.
Our starting dataframes are too ideal, so we’ll begin by spliting them to make it more realistic for an email campaign. We’ll take the leading 60 rows of the email_content, then ignore the 1st 10 email_id values from user_email.
This results in missing keys for email_id at the start, and id at the end
The left contains values not in the right, and the right contains values not in the left (Source: SO)
We can use
to establish the same relationships from our previous section, completing an inner and outer join as an intersection or union. Merge is quite useful when their columns’ names don’t match but we still want to relate them together. In this case our matching keys are the id in the campaign and email_id of the inbox.
To perform an intersection, we need 2 matching keys, and pass
. To check our result by an additional parameter called
and set it to
Areas where email_id and id have matching values 11 to 60.
This is the correct result as the only valid
value for an intersection is
To perform a union, we repeat the process of matching keys, and pass
This time we’ll switch the order of the dataframes, choosing campaign on the left and inbox on the right. The results show the columns in the merge are displayed differently.
Include all values in both, id and email_id ranging from 1 to 100.
can do more than
, it can also perform the relative joins. This is categorized as a left or right join and may be used to relate the datasets. One such relation is columns, or features for ML models. Another relation is to introduce more rows, or samples for ML models.
To start off, let’s do a left join on the user_inbox dataframe with the email_content.
Since email content and user inbox share the email_id we choose that as our parameter.
For our matching values, we have id from email_content, and on the left we have email_id from user_inbox. We’ll add back the indicator to check our result.
From the campaign dataset there are values for ids ranging from 1 to 60.
As a check, note that the
column doesn’t contain values for
To create the right join in Pandas, we
and pass the “on” values, using how=’right’. We include the indicator to check once again.
From the inbox dataset, the email_id ranges from 11 to 100.
A common misconception is that if we flip the order of the datasets, it will be the same as switching from left to right. This is wrong, because choosing which dataset is on the right and which dataset is on the left decides the column order as well.
Here we switch inbox and email, and perform another intersection.
The dimensions match the 1st example, but the order of the columns are changed.
If you haven’t noticed, there’s the 2 pesky identical columns problem. In part 1, our previous approach was to convert the dataframes using
but here we don’t have a shared key. We can accomplish removing identical columns by applying the
method. This is a much better alternative, as its built-in functionality accomplishes it faster, looks cleaner, and with minimal effort or code.
Before we can call
, we need to tell it what the similarities are. By default,
uses the index to determine which values to join on. As a result, the set index function is used similarly to merge’s “left_on=”, and “right_on=”. By changing the column into an index, it removes the column entirely. This is reminiscent of using drop_duplicates. But if you want the columns to remain you can include ‘drop=False’ to keep the columns.
Join works just like merge, instead of providing a list, the caller will represent the left, and the argument will represent the right,
is used to specify the match.
Notice how the index columns for id and email_id are gone now
With drop=False, the result is similar to merge.
Result is similar, but the index values differ. Merge keeps row number, join was overwritten.
does, but each excels at different situations. In most cases,
is more versatile and therefore used more. In the next part, we’ll apply what we’ve learned to handle larger datasets, thus completing the entire marketing campaign featuring data on emails, the inbox, and user statistics.