ION: a pertinent new measure for mining information from many types of data

Since last decade, many methods with appropriate measures are proposed in knowledge discovery in databases. These measures aim at both improving the quality of mined association rules and reducing the problem of many nested rules. This paper presents a new statistical Implication Oriented Normalized measure, denoted ION. ION turns to be a unifying framework for several probabilistic measures of interestingness of association rules mined from diverse kind of dataset. It naturally leads to a pertinent algorithm for mining statistical implication, according to logical reasoning: one has the identity ION(¬b→¬a) ≡ ION(a → b), for any itemsets a and b. In addition, it takes into account both of positively or negatively oriented dependencies and of a deviation from equilibrium on large databases.