AAAI Publications, Twenty-Third International FLAIRS Conference

Font Size: 
Mining Actionable Patterns
Prabakararaj Swapna Raj, Ravindran Balaraman

Last modified: 2010-05-06


The goal of data mining is to mine interesting patterns that are actionable, i.e., support decision making. Various measures of interestingness that have been considered in the literature, such as classification accuracy, support etc., are based on intrinsic statistical properties of the data. Data mining algorithms typically operate by optimizing such measures. While researchers have looked at measuring the utility of discovered patterns in decision making, there has not been much work on using such utility measures for driving the mining process as such. In our work we propose a generic framework for ``closing the loop", i.e., using utility in decision making to drive the mining process. We use concepts from meta-learning in order to tune the hyper-parameters of mining algorithm thus influencing the patterns detected. We build on earlier work by Elovici and Braha, that uses decision theory for formulating an utility measure, to specialize the framework for classification tasks. We show empirical validation of the approach on a simple test domain.


Data Mining, Meta-learning

Full Text: PDF