Comparing Expert and Metric-Based Assessments of Association Rule Interestingness
Diego Luna-Bazaldua, Ryan S. Baker, Maria Ofelia San Pedro · Educational Data Mining · 2014
In association rule mining, interestingness refers to metrics that are applied to select association rules, beyond support and confidence. For example, Merceron & Yacef (2008) recommend that researchers use a combination of lift and cosine to select association rules, after first filtering out rules with low support and confidence. However, the empirical basis for considering these specific metrics to be evidence of interestingness is rather weak. In this study, we examine these metrics by distilling association rules from real educational data relevant to established research questions in the areas of affect and disengagenment. We then ask three domain experts to rate the interestingness of the resultant rules. We finally analyze the data to determine which metric(s) best agree with expert judgments of interestingness. We find that Merceron & Yacef (2008) were right. Lift and cosine are good indicators of interestingness. In addition, the Phi Coefficient, Convinction, and Jaccard also turn out to be good indicators of interestingness.