Frequent Item Set Mining and Association Rules

One of the important categories of data mining problems is that of associations between attributes. This gives useful insight for such diverse business problems as product cross-selling, website perception, and decision problems. There are two ways to look at attribute associations. The first is on the attribute-level, i.e., one looks for statistical dependencies between the attributes. Graphical models are a tool for such associations. The second way is at the value-level. Association rules [2] are the premier tool for this class of problems. Informally, association rules can be seen as a kind of if-then rules: if a person buys a newspaper, he or she also buys chocolate. The twist association rules bring to classical if-then rules is a conditional probability. If a person buys a newspaper, there is a probability that he or she will also buy chocolate. This conditional probability is known as confidence in the literature. Another measure that is generally associated with association rules is that of support : the fraction of customers for whom the rule holds, or rather the relative number of customers buying all items occurring in the rule (the socalled underlying itemset). If there is just one customer that buys newspapers and he or she also happens to buy chocolate, the association rule is not very interesting. Next to being an interestingness measure, the support of a rule also plays a key role in the standard algorithms for association rule discovery. Given thresholds minsup and minconf for the support and confidence, these algorithms compute all association rules whose support and confidence exceed these thresholds. Itemsets with support at least minsup are called frequent. Originally, association rules were introduced in a binary, non-temporal setting. For example, one considered a collection of transactions at the check-out of a store only recording whether a certain item was bought or not. Later, many extensions have been introduced, e.g., sequences (time), hierarchical clusters of items, and item-counts. Also more complex patterns then itemsets, for example trees and graphs, have been considered extensively.