Probabilistic Measures for Interestingness of Deviations - A Survey
Adnan Masood, Sofiane Ouaguenouni · International Journal of Artificial Intelligence & Applications · 2013
Association rule mining has long being plagued with the problem of finding meaningful, actionable knowledge from the large set of rules.In this age of data deluge with modern computing capabilities, we gather, distribute, and store information in vast amounts from diverse data sources.With such data profusion, the core knowledge discovery problem becomes efficient data retrieval rather than simply finding heaps of information.The most common approach is to employ measures of rule interestingness to filter the results of the association rule generation process.However, study of literature suggests that interestingness is difficult to define quantitatively and can be best summarized as, a record or pattern is interesting if it suggests a change in an established model.Almost twenty years ago, Gregory Piatetsky-Shapiro and Christopher J. Matheus, in their paper, "The Interestingness of Deviations," argued that deviations should be grouped together in a finding and that the interestingness of a finding is the estimated benefit from a possible action connected to it.Since then, this field has progressed and new data mining techniques have been introduced to address the subjective, objective, and semantic interestingness measures.In this brief survey, we review the current state of literature around interestingness of deviations, i.e. outliers with specific interest around probabilistic measures using Bayesian belief networks.