Selecting Relevant Association Rules From Imperfect Data

Abstract

Association Rule Mining (ARM) in the context of imperfect data (e.g. imprecise data) has received little attention so far despite the prevalence of such data in a wide range of real-world applications. In this work, we present an ARM approach that can be used to handle imprecise data and derive imprecise rules. Based on evidence theory and Multiple Criteria Decision Analysis, the proposed approach relies on a selection procedure for identifying the most relevant rules while considering information characterizing their interestingness. The several measures of interestingness defined for comparing the rules as well as the selection procedure are presented. We also show how a priori knowledge about attribute values defined into domain taxonomies can be used to (i) ease the mining process, and to (ii) help identifying relevant rules for a domain of interest. Our approach is illustrated using a concrete simplified case study related to humanitarian projects analysis.