Classification based on Association Rules: Complexity and Interestingness Guided Algorithm

Abstract

Classification based on association rules has proved able to get high precision and deal with data noise in real scenarios. However, very low support thresholds are used to get these results, generating a great number of rules. In this process, interesting and non-interesting itemsets are considered for subsequent rule generation. In this paper, we propose a new method to identify not only frequent but also interesting itemsets, including an interestingness measure when pruning candidate itemsets. With this approach, we reduce the number of association rules in the final classifier and increase precision. We combine it with an iterative approach, reducing the execution time and complexity of the final rule set. We have tested the proposal using a COVID19 patient database.