Data Stream Classification Framework for Data Streams using Dynamic Ensemble Selection

With the progressive advancement in the area of information technology, an huge amount of data is produced and transmitted every second in the real world. To address imbalanced data with idea drift for data streams, a novel framework for data stream classification is proposed. The class imbalance problem in data stream is handled using the proposed Adaptive Synthetic Oversampling (ASYNO) algorithm and if the concept drift occurs in data streams, then there is a possibility of sudden decrease in the balanced accuracy score. A novel dynamic classification framework based on the Dynamic Ensemble Selection (DES) was proposed in order to address the deteriorated balanced accuracy score and geometric mean in the incoming data stream. The classifiers which classify minimum one example from the competence region is chosen by KNORA-U. The number of votes obtained from each classifier is equal to the number of instances in its competence region for which, it predicts the target class. In case to arrive at the last ensemble decisionmaking, the votes received by each dynamically selected base classifiers are merged using majority voting. The proposed framework using DES, provides high balanced accuracy of 98.4% and geometric mean of 99.4% when compared to existing methods used for balancing the classes and for handling concept drift in data streams.