Features and Classes Drift Detector to Deal with Imbalanced Data Streams

Data streams, due to their dynamic nature, tend to impose a number of constraints on the functioning of the learning models used to extract knowledge from these environments. In this context, concept drift is an emerging research area, as they negatively affect the performance of classifiers: after they have been trained with a specific concept, they tend to lose accuracy in the presence of a new concept. Additionally, this problem is often worsened in environments with imbalanced classes, because the identification of changes in the distributions of examples belonging to minority classes is usually more complex, due to their lack of representativeness in the data stream. This work proposes the Features and Classes Drift Detector (FCDD), a new method specially designed to deal with the problem of concept drifts in imbalanced data streams, aiming to maintain the accuracy of detections in minority classes and, in addition, to avoid discarding the knowledge inherent to the classes unaffected by drifts. Experiments conducted in an imbalanced scenario with partial concept drift demonstrated the effectiveness of the proposed method when compared to the current state of the art detectors.