A novel algorithm weighting different importance of classes in enhanced association rules

Abstract

Importance of methods for explainable AI has been ever growing within many processes; a simple form of knowledge is needed by humans who operate the processes. One of the methods suitable for explainable AI is represented by enhanced association rules on categorical data. This method admits an immediate interpretation. A problem arises when we are interested in finding circumstances for a rare category. Even when we find such a category (e.g., in a subset where this category appears five times more), the overall share of this category may be too small for using this rule within processes (e.g., sending someone to an expensive examination, performing calls to clients), and may therefore be refused by process executors (doctors, salesmen). That is why we have created a new algorithm based on association rule mining called UIC-Miner (UIC stands for Unequal Importance of Categories), which overcomes this problem. This algorithm has a pilot implementation for the purposes of testing and reproducing. The algorithm itself uses categories weighted by their levels of importance. This paper first describes a typical situation when category weights are useful. Then, the new algorithm is described, and several examples of its applications are presented. Finally, possible applications of the available tools are discussed.