Otmização de Classificadores com Mineração de Padrões Excepcionais
Traditional predictive algorithms look for optimizing global predictive metrics in order to identify a tendency which describes the dataset behavior. Then, this generalized rule can be used to provide insights on unseen samples. This modus operandi may lead to an overlook of interesting patterns that are not close to the main tendency which contains relevant information. These potentially interesting regions in the space are called local exceptions, and can be identified by algorithms such as the Exceptional Model Mining (EMM) in the form of subgroups. The EMM generalizes the subgroup discovery task by covering also multi-variables quality measures. Then, its instances are defined by the combination of a model class over target attributes and a quality measure over the model class. Explicitly adding this information as new features into a dataset would force classifiers to consider these local exceptionalities into their search space, supporting decision boundaries building process and, consequentially, improving predictive performance. Our results suggest that, by adding the subgroups found with EMM as new features into a dataset it was possible to improve the predictive performance of traditional linear classifiers when compared with the same algorithms, running with the same parameters over the same dataset but, without the EMM features.