Mining action rules from scratch

Action rules provide hints to a business user what actions (i.e., changes within some values of flexible attributes) should be taken to improve the profitability of customers. That is, taking some actions to re-classify some customers from less desired decision class to the more desired one. However, in previous work, each action rule was constructed from two rules, extracted earlier, defining different profitability classes. In this paper, we make a first step towards formally introducing the problem of mining action rules from scratch and present formal definitions. In contrast to previous work, our formulation provides guarantee on verifying completeness and correctness of discovered action rules. In addition to formulating the problem from an inductive learning viewpoint, we provide theoretical analysis on the complexities of the problem and its variations. Furthermore, we present efficient algorithms for mining action rules from scratch. In an experimental study we demonstrate the usefulness of our techniques.