Data Analysis From Scratch With Python: Step By Step Guide

13. Association Rule Learning
This is a continuation of Unsupervised Learning. In the previous chapter we’ve
discovered natural patterns and aggregates in Mall_Customers.csv. There was
not much supervision and guidance on how the “correct answers” should look
like. We’ve allowed the algorithms to discover and study the data. As a result,
we’re able to gain insights from the data that we can use.
In this chapter we’ll focus on Association Rule Learning. The goal here is
discover how items are “related” or associated with one another. This can be
very useful in determining which products should be placed together in grocery
stores. For instance, many customers might always be buying bread and milk
together. We can then rearrange some shelves and products so the bread and milk
will be near to each other.
This can also be a good way to recommend related products to customers. For
example, many customers might be buying diapers online and then purchasing
books about parenting later. These two products have strong associations
because they mark the customer’s life transition (having a baby). Also if we
notice a demand surge in diapers, we might also get ready with parenting books.
This is a good way to somehow forecast and prepare for future demands by
buying supplies in advance.
In grocery shopping or any business involved in retail and wholesale
transactions, Association Rule Learning can be very useful in optimization
(encouraging customers to buy more products) and matching supply with
demand (e.g. sales improvement in one product also signals the same thing to
another related product).
Explanation
So how do we determine the “level of relatedness” of items to one another and
create useful groups out of it.? One straightforward approach is by counting the
transactions that involve a particular set. For example, we have the following
transactions:
Transaction
Purchases
1
Egg, ham, hotdog

---

2
Egg, ham, milk
3
Egg, apple, onion
4
Beer, milk, juice
Our target set is {Egg, ham}. Notice that this combination of purchases occurred
in 2 transactions (Transactions 1 and 2). In other words, this combination
happened 50% of the time. It’s a simple example but if we’re studying 10,000
transactions and 50% is still the case, of course there’s a strong association
between egg and ham.
We might then realize that it’s worthwhile to put eggs and hams together (or
offer them in a bundle) to make our customers’ lives easier (while we also make
more sales). The higher the percentage of our target set in the total transactions,
the better. Or, if the percentage still falls under our arbitrary threshold (e.g. 30%,
20%), we could still pay attention to a particular set and make adjustments to our
products and offers.
Aside from calculating the actual percentage, another way to know how
“popular” an itemset is by working on probabilities. For example, how likely is
product X to appear with product Y? If there’s a high probability, we can
somehow say that the two products are closely related.
Those are ways of estimating the “relatedness” or level of association between
two products. One or a combination of approaches might be already enough for
certain applications. Perhaps working on probabilities yields better results. Or,
prioritising a very popular itemset (high percentage of occurrence) results to
more transactions.
In the end, it might be about testing different approaches (and combinations of
products) and then seeing which one yields the optimal results. It might be even
the case that a combination of two products with very low relatedness allow for
more purchases to happen.
Apriori
Whichever is the case, let’s explore how it all applies to the real world. Let’s call
the problem “Market Basket Optimization.” Our goal here is to generate a list of

---

sets (product sets) and their corresponding level of relatedness or support to one
another. Here’s a peek of the dataset to give you a better idea: