Adaptive Windowing (ADWIN3) to Learning from Time-Changing Data Stream

Classification of data streams has become an important and active field of study. The primary characteristics of data streams are a large quantity of incoming data, rapid arrival rate, and changes in their nature and distribution over time. Data Streams are the result of diverse high speed networked devices, in which data trends, seasonality, and drift detections are the greatest issues in classification tasks. In this study, we proposed a technique for designing algorithms that are capable of adapting to changing data streams. We employed the Hoeffding_Adaptive_Tree (HAT) as the foundation of our new approach in addition to EDDM. The proposed system is called ADaptive_WINdowing_3 (ADWIN3), which can handle dispersion and concept drift. To extend these assurances to the resulting adaptive learning process, which is based on the strategic placement of the change detector and estimator modules, we selected implementations that come with theoretical guarantees. Our technique has the major advantage of not requiring any assumptions regarding the rate or frequency of stream changes, whereas other systems often include many user_defined parameters. In our testing, the novel techniques outperformed HAT, A popular technique for tree induction in drifting data streams.