A SVR-based ensemble approach for drifting data streams with recurring patterns

Pattern drift is a common issue for machine learning in real applications, as the distribution generating the data may change under nonstationary environmental/operational conditions. In our previous work, a strategy based on Feature Vector Selection (FVS) has been proposed for enabling a Support Vector Regression (SVR) model to adaptively update with streaming data, but the proposed strategy suffers from the incapability of treating recurring patterns. An instance-based online learning approach is proposed in this paper, which can adaptively update an SVR-based ensemble model with steaming data points. The proposed approach reduces the computational complexity of the updating process by selecting only part of the newly available data and allows following timely the ongoing patterns by resorting to FVS. The proposed approach creates new sub-models directly from a basic model and the sub-models represent separately the data stream at different periods. A dynamic ensemble selection strategy is integrated in the approach to select the sub-models most relevant to the new data point for deriving the prediction, while reducing the influence of the irrelevant ones. The weights of the different models in the ensemble are updated, based on their prediction errors. Comparison results with several benchmark approaches on several synthetic datasets and on the dataset concerning the leakage from the first seal in a Reactor Coolant Pump, prove the efficiency and accuracy of the proposed online learning ensemble approach.