Look at these two rules:
- {Oatmeal, corn flakes → Milk}
- {Dog food, paperclips → Washing powder}
Intuitively, the second rule looks more unlikely than the first one, doesn't it? How can we tell that for sure, though? In this case, we need some quantitative measures that will show us how likely each rule is. What we are looking for here are association measures, as we call them in machine learning and data mining. Rule mining algorithms revolve around this notion in a similar manner to how distance-based algorithms revolve around distance metrics. In this chapter, we're going to use four association measures: support, confidence, lift, and conviction (see Table 5.1).
Note that these measures tell us nothing about how useful or interesting the rules are, but only quantify their probabilistic characteristics. A rule's usefulness and practicality can be hard to grasp mathematically and often requires human judgment in each case. As usual in statistics, interpreting...