Interestingness Measures for Classification Rule Mining: Model Selection Ability
This study analyzes the ability of interestingness measures to capture the correct predictive model through a new simulation design. The simulation is designed to be general enough to allow fair conclusions to be drawn without depending on subjective opinions. We found that the relative success of interestingness measures in capturing the correct model depends on two major factors: (1) the characteristics of the model and (2) the weight of each of three components—confidence, support, and the probability of result or class—in the measures formula. No measure was found to perform best in all scenarios. However, we found that two groups of measures work well with different scenarios and that these groups of measures complement each other. That is, in any given scenario, the measures in one or the other of these two groups would perform the best. Therefore, in actual practice, when the characteristics of data are unknown, we propose using representatives of these two groups—confidence and lift—to capture the predictive models as either of these will capture the predictive model that best fits the given scenario.
Keywords: Classification rule mining, Interestingness measures, Predictive model
Download Full-Text
ABOUT THE AUTHORS
Pannapa Changpetch
Bentley University
Dennis K. J. Lin
Penn State University
Pannapa Changpetch
Bentley University
Dennis K. J. Lin
Penn State University
