Adaptive stochastic configuration network based on online active learning for evolving data streams

Abstract

Stochastic Configuration Networks (SCNs) have exhibited significant potential in data mining, owing to their advantages in fast incremental construction and universal approximation capabilities. However, less researches were done on SCNs-based classification models for concept-drifting data streams. The so-called drifts refer to data distributions changing over time that may degrade the classification performance of SCNs trained on historical data. The previous drift adaptation approach is to discard all the hidden nodes of SCNs, and then learn a new model with new instances, in which the valuable historical information cannot be fully utilized. In addition, labeling all newly-arrived instances is time-consuming and impractical. To address these issues, an adaptive stochastic configuration network embedding online active learning is proposed. Crucially, a query strategy is developed to select representative instances for labeling based on the change degree of instances density and their uncertainty. An online update mechanism is employed to incrementally update the network's output parameters instance by instance. To rationally forget the outdated information and learn new concepts, a dynamic adjustment mechanism adaptively adds or prunes nodes in the SCN model. Experimental results for nine datasets confirm that our algorithm outperforms six popular ones on classification accuracy.