Affiliations: Department of Computer Science, University of Houston, Houston, TX, USA
Correspondence:
[*]
Corresponding author: C. Ordonez, University of Houston, Department of Computer Science, Houston, TX 77204, USA. E-mail: ordonez@cs.uh.edu.
Abstract: Association rules and decision trees represent two well-known data mining techniques to find predictive rules. In this work, we present a detailed comparison between constrained association rules and decision trees to predict multiple target attributes. We identify important differences between both techniques for such goal. We conduct an extensive experimental evaluation on a real medical data set to mine rules predicting disease on multiple heart arteries. The antecedent of association rules contains medical measurements and patient risk factors, whereas the consequent refers to the degree of disease on one artery or multiple arteries. Predictive rules found by constrained association rule mining are more abundant and have higher reliability than predictive rules induced by decision trees. We investigate why decision trees miss certain rules, why they tend to have lower confidence and the possibility of improving them to match constrained association rules. Based on our experimental results, we show association rules, compared to decision trees, tend to have higher confidence, they involve larger subsets of the data set, they are better for multiple target attributes, they work better with user-defined binning and they are easier to interpret.
Keywords: Association rule, decision tree, classification, search constraint
