Selecting structural patterns for classification

Many techniques have recently been proposed for discovering structural patterns. Using the discovered structural patterns as features for classification has shown success in some application domains. However, the efficiency and effectiveness of such a classification algorithm is often impeded by the huge number of structural patterns discovered by the associated structural pattern mining algorithm. In this paper, we focus on the feature selection problem of structural patterns. The goal is to develop a scheme that effectively selects a subset of structural patterns as the features for the following induction algorithm. We show how to make use of the downward closure property inherent in the structural patterns to design a novel feature selection algorithm. We also evaluate our algorithm by applying the real-world health insurance data for building a classification model to detect health care fraud and abuse. The experimental results show that a great extent of redundant features can be eliminated by our feature selection algorithm, resulting in both accuracy improvement and computation cost reduction.