An Empirical Study of Online Learning in Non-stationary Data Streams Using Ensemble of Ensembles

Numerous information system applications produce a huge amount of non-stationary streaming data that demand real-time analytics. Classification of data streams engages supervised models to learn from a continuous infinite flow of labeled observations. The critical issue of such learning models is to handle dynamicity in data streams where the data instances undergo distributional change called concept drift. The online learning approach is essential to cater to learning in the streaming environment as the learning model is built and functional without the complete data for training in the beginning. Also, the ensemble learning method has proven to be successful in responding to evolving data streams. A multiple learner scheme boosts a single learner's prediction by integrating multiple base learners that outperform each independent learner. The proposed algorithm EoE (Ensemble of Ensembles) is an integration of ten seminal ensembles. It employs online learning with the majority voting to deal with the binary classification of non-stationary data streams. Utilizing the learning capabilities of individual sub ensembles and overcoming their limitations as an individual learner, the EoE makes a better prediction than that of its sub ensembles. The current communication empirically and statistically analyses the performance of the EoE on different figures of merits like accuracy, sensitivity, specificity, G-mean, precision, F1-measure, balanced accuracy, and overall performance measure when tested on a variety of real and synthetic datasets. The experimental results claim that the EoE algorithm outperforms its state-of-the-art independent sub ensembles in classifying non-stationary data streams