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 join and merge.
- Before we begin
- Merge Dataframes
- Join Dataframes
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 join, merge, or concat. 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 merge and join 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.
In this guide, we’ll reuse our email_content dataset and include the user_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 to part 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 introductory guide before continuing.
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.
In Pandas, the merge 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.
We can use merge 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 the 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 how=’inner’. To check our result by an additional parameter called indicator and set it to True.
This is the correct result as the only valid _merge value for an intersection is both.
To perform a union, we repeat the process of matching keys, and pass how=‘outer’.
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.
But merge can do more than concat, 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.
As a check, note that the _merge column doesn’t contain values for right_only.
To create the right join in Pandas, we merge and pass the “on” values, using how=’right’. We include the indicator to check once again.
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.
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 concat and drop_duplicates, but here we don’t have a shared key. We can accomplish removing identical columns by applying the join and set_index 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 join, we need to tell it what the similarities are. By default, join 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, set_index is used to specify the match.
With drop=False, the result is similar to merge.
Merge handles everything join and concat does, but each excels at different situations. In most cases, merge 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.