Adaptive Bagging Methods for Classification of Data Streams with Concept Drift
Martin Sarnovský, Jan Marcinko · Acta Polytechnica Hungarica · 2021
Data streams represent a continuous stream of data, in many forms, coming from different sources.Streams are often dynamic and its underlying structure usually changes over time.When solving predictive tasks on the streaming data, traditional models, trained on historical data, may become invalid, when such change occurs.Therefore, adaptive models, equipped with mechanisms to reflect the changes in the data, are suitable to solve these tasks.Adaptive ensemble models represent a popular group of such methods used in classification tasks on data streams.In this paper, we designed and implemented the modifications of the adaptive bagging methods, which utilizes internal class-weighting schemes for the model adaptation.Implemented models were evaluated on two simulated real-world data streams and compared with base classifiers and other adaptive methods.In addition to the performance evaluation, we also analyzed other models' characteristics, such as the duration of model update and memory requirements.