Chapter 2. Association Rule Mining

In our data mining toolbox, measuring the frequency of a pattern is a critical task. In some cases, more frequently occurring patterns may end up being more important patterns. If we can find frequently occurring pairs of items, or triples of items, those may be even more interesting.

In this chapter, we begin our exploration of frequent itemsets, and then we extend those to a type of pattern called association rules. We will cover the following topics:

• What is a frequent itemset? What are the techniques for finding frequent itemsets? Where are the bottlenecks and how can we speed up the process?

• How can we extend a frequent itemset to become an association rule?

• What makes a good association rule? We will learn to describe the value of a particular association rule, given its level of support in the database, our confidence in the rule itself, and the value added by the rule we found.

To do this, we will write a program to find frequent itemsets in an open...

What are frequent itemsets?

Finding frequent itemsets is a type of counting activity. But unlike producing a simple tally of items we observe in a dataset (today we sold 80 carrots and 100 tomatoes), finding frequent itemsets is slightly different. Specifically, to find frequent itemsets we look for co-occurring sets of items within some larger group. These larger groups are sometimes imagined as supermarket transactions or shopping baskets, and the entire exercise is sometimes called market basket analysis. Staying with the supermarket analogy, the items co-occurring within those baskets are sometimes imagined to be combinations of products purchased at the supermarket. For example, given a set of supermarket transactions or baskets, we might be interested in whether the itemset of {carrots, tomatoes} occurs more frequently in baskets than does the {cucumbers, lemons} itemset.

The purpose of frequent itemset mining is to make interesting discoveries of co-occurring items within a set of...

Towards association rules

All of this frequent itemset stuff is fine, but we are ultimately on the hunt for association rules, which are much more exciting. Association rules are formed from frequent itemsets, with a few small twists. We are interested in making a statement about the frequent itemsets like this: people who buy vanilla wafers also buy bananas 60% of the time. In order to do so, we need to learn how to calculate a few additional metrics, starting with two we call support and confidence.

Support

If we are looking for frequent itemsets, then we also need a way to express how often we see these sets occurring in baskets, and whether that number qualifies as frequent. If I see {vanilla wafers, bananas} in 90% of baskets, is that considered frequent? What about 50% of baskets? What about 5%? We call this number the support of the itemset. The support is just the number of times we saw that itemset over all the baskets.

To make support more meaningful, and to begin talking about "interestingness...

A project – discovering association rules in software project tags

In 1997, the website, Freshmeat, was created as a directory that tracked free, libre, and open source software (FLOSS) projects. In 2011, the site was renamed Freecode. After sales and acquisitions and several site redesigns, in 2014 all updates to the Freecode site were discontinued. The site remains online, but it is no longer being updated and no new projects are being added to the directory. Freecode now serves as a snapshot of facts about FLOSS projects during the late 1990s and 2000s. These facts about each software project include its name, its description, the URL to download the software, tags that describe its features, a numeric representation of its popularity, and so on.

As part of my FLOSSmole project, I have catalogued data from Freshmeat/Freecode since 2005. Freshmeat/Freecode provided periodic RDF downloads describing each project on the site. I downloaded these, parsed out the project data, organized it into...

Summary

In this chapter, we learned how to generate frequent itemsets from a dataset using the Apriori algorithm. We then proposed association rules from these itemsets by describing their support and confidence. We used one additional check, an added value measure, to ensure that the proposed rules were interesting. We implemented all these concepts using a freely available dataset of Freecode open source projects and their tags. We calculated support for single tags, then generated doubletons and tripletons that met a minimum support threshold. For rules with one item on the right-hand side, we calculated confidence and added value for each. Finally, we looked closely at the rules that were generated and tried to figure out which ones were interesting, using the metrics we had calculated.

In the next chapter, we will continue our quest to make connections between items in a data set. However, unlike in this chapter where we were trying to find groups of two or three items that are already...