Action Rule Mining

This dissertation focuses on the research and development of methods for action rule mining, which is an extension of standard classification rules. Classification rules are used to predict the value of a target attribute based on the values of other attributes for a given instance. Action rules additionally provide information on what changes to attribute values should be made to improve the classification toward achieving a desired goal. This approach enables the recommendation of changes and the generation of counterfactual (what-if) explanations, exploring what changes would lead to the desired outcome. The Action-Apriori algorithm, developed as part of this work, represents a modification of the Apriori algorithm specifically tailored for action rule mining. This algorithm facilitates efficient rule generation by integrating user-defined parameters into the rule creation process, thus reducing the number of candidate combinations of attribute values and computational overhead. An enhancement introduced is the uplift metric, which evaluates the effectiveness of proposed actions, alongside the concept of dominant action rules, which filters and retains only the most impactful rules. Subsequently, support for high-utility action rule mining was added, allowing for the consideration of costs and benefits when recommending changes. As part of this dissertation, two open-source packages for action rule mining were developed. The actionrules-lukassykora package focuses on the implementation of existing algorithms for action rule mining, while the action-rules package employs the author’s Action-Apriori algorithm and also leverages GPU acceleration to further enhance its computational performance. This work additionally compares both packages, evaluating differences in usability and performance.