Rules, Subgroups and Redescriptions as Features in Classification Tasks

We evaluate the suitability of using supervised and unsupervised rules, subgroups and redescriptions as new features and meaningful, interpretable representations for classification tasks. Although using supervised rules as features is known to allow increase in performance of classification algorithms, advantages of using unsupervised rules, subgroups, redescriptions and in particular their synergy with rules are still largely unexplored for classification tasks. To research this topic, we developed a fully automated framework for feature construction, selection and testing called DAFNE - Descriptive Automated Feature Construction and Evaluation. As with other available tools for rule-based feature construction, DAFNE provides fully interpretable features with in-depth knowledge about the studied domain problem. The performed results show that DAFNE is capable of producing provably useful features that increase overall predictive performance of different classification algorithms on a set of different classification datasets.