Informativity of Association Rules from the Viewpoint of Information Theory
Dmytro Sytnikov, O. Titova, Sergii Minukhin, Andriy Kovalenko, Serhii Titov · 2018
In this paper information theory, in particular, mutual information is suggested as a possible basis for association rule quality assessment. Normally association rules are not considered to be associated with mutual information, although there are a lot of papers on the matter of association significance. We have introduced a new approach supported by precise mathematical formulas, which allows considering non-binary associations in the context of Shannon's theory. An overview of some parameters important for determining the importance of an association dependence has been presented. Complete mutual information that can be calculated as the sum of four addends has been investigated in the context of various requirements to association rule importance. It has been shown which of the requirements are met by the complete mutual information and which ones are met with one of the addends used to calculate it. The obtained results can be used for assessing association rules discovered in big data from the viewpoint of information theory.