Rival Learner Algorithm with Drift Adaptation for Online Data Stream Regression
Zhenwei Liao, Yongheng Wang · 2018
Real-time extraction of meaningful data streams patterns is an increasingly important issue for machine learning and data mining communities. In this paper, we proposed a regression algorithm incremental on data streams which are infinite, high-speed and time-varying. The algorithm integrates two incremental model trees, global and local models. In order to cope with the concept drift, we compare a global sub-model with a local sub-model. A global sub-model is trained based on its historical sample, while the local sub-model is trained based on its sliding window. Although the method uses a local sub-model as a drift indicator, it uses a global sub-model to predict. The model is executed in real time online, and each example is observed only once at the speed of arrival and at any time to maintain a readily available model. Moreover, because of frequent to retrain model (local sub-model alternative global sub-model), makes the complexity of the model tree reduced. The algorithm has a drift detection adaptation mechanism, which can maintain accurate and updated regression models at any time. This method improves performance, greatly reduces errors, whether in stationary or non-stationary data streams, and greatly reduces the adaptation cost.