Probabilistic Frequent Itemset Mining with Hierarchical Background Knowledge

Abstract

Abstract-In the recent years, there has been significant development in the field of Probabilistic Frequent Itemset Mining (PFIM). Despite the complexity of calculating the frequentness probability of an itemset, approximation techniques allow us to reduce the complexity of the problem with very low approximation error. In this paper we investigate how to incorporate hierarchical taxonomies into the attribute uncertainty model, which assumes independence between the existential probability of items in a transaction. We propose scalable methods which can reduce noise, and ensure consistency of the transactions by approximating the dependencies between attributes implied by a background hierarchical taxonomy. We also perform experiments in order to evaluate the scalability, accuracy of the approximation, as well as the denoising performance of the proposed methods.Index Terms-Probabilistic frequent itemset mining, generalized rules, hierarchical background knowledge.