View article

Educational data differs from traditional knowledge discovery domains in several ways. One of them is the fact that it is difficult, or even impossible, to compare different methods or measures a posteriori and deduce which the best is. It is therefore essential to use techniques and measurements that are fairly intuitive and easy to interpret. Extracting the most interesting association rules can be quite tricky. One of the difficulties is that many measures of interestingness do not work effectively for all datasets and are hard to understand intuitively by the teachers. We argue in this paper that cosine and added value (or equivalently lift) are well suited to educational data, and that teachers can interpret their results easily. We argue that interestingness should be checked with cosine first, and then with lift if cosine rates the rule as noninteresting. If both measures disagree, teachers should use the intuition behind the measures to decide whether or not to dismiss the association rule. We provide a case study with data from a LMS.