321 Pasquier

In the domain of knowledge discovery in databases and its computational part called data mining, many works addressed the problem of association rule extraction that aims at discovering relationships between sets of items (binary attributes). An example association rule fitting in the context of market basket data analysis is cereal ∧ milk → sugar (support 10%, confidence 60%). This rule states that 60% of customers who buy cereals and sugar also buy milk, and that 10% of all customers buy all three items. When an association rule support and confidence exceed some user-defined thresholds, the rule is considered relevant to support decision making. Association rule extraction has proved useful to analyze large databases in a wide range of domains, such as marketing decision support; diagnosis and medical research support; telecommunication process improvement; Web site management and profiling; spatial, geographical, and statistical data analysis; and so forth. The first phase of association rule extraction is the data selection from data sources and the generation of the data mining context that is a triplet D = (O, I, R), where O and I are finite sets of objects and items respectively, and R ⊆ O × I is a binary relation. An item is most often an attribute value or an interval of attribute values. Each couple (o, i) ∈ R denotes the fact that the object o ∈ O is related to the item i ∈ I. If an object o is in relation with all items of an itemset I (a set of items) we say that o contains I. This phase helps to improve the extraction efficiency and enables the treatment of all kinds of data, often mixed in operational databases, with the same algorithm. Datamining contexts are large relations that do not fit in main memory and must be stored in secondary memory. Consequently, each context scan is very time consuming.