Evaluation of Position-Constrained Association-Rule-Based Classification for Tree-Structured Data

Abstract

Tree-structured data is popular in many domains making structural classification an important task. In this paper, a recently proposed structure preserving flat representation is used to generate association rules using itemset mining techniques. The main difference to traditional techniques is that subtrees are constrained by the position in the original tree, and initial associations prior to subtree reconstruction can be based on disconnected subtrees. Imposing the positional constraint on subtreee typically result in a reduces the number of rules generated, especially with greater structural variation among tree instances. This outcome would be desired in the current status of frequent pattern mining, where excessive patterns hinder the practical use of results. However, the question remains whether this reduction comes at a high cost in accuracy and coverage rate reduction. We explore this aspect and compare the approach with a structural classifier based on same subtree type, but not positional constrained in any way. The experiments using publicly available real-world data reveal important differences between the methods and implications when frequent candidate subtrees on which the association rules are based, are not only equivalent structure and node label wise, but also occur at the same position across the tree instances in the database.