A Semi-Supervised Active Learning Neural Network for Data Streams With Concept Drift

Learning from data streams originating from non-stationary environments is vital for many real-world applications. A notable challenge in this task is concept drift. Most existing methods rely on a large number of labeled instances to detect and tackle concept drift in data streams. However, obtaining labeled data is not easy due to high costs, especially in potentially infinite data streams. To address this issue, a semi-supervised active learning neural network for data streams with concept drift is proposed to build an accurate classification model from incompletely labeled data. First, a semi-supervised regularization based on smoothness assumption is proposed to adjust the network weights and utilize unlabeled data. Second, a smoothness loss-based query strategy is designed to select instances that effectively improve model performance in line with the semi-supervised regularization objective. Notably, the query strategy and semi-supervised regularization form a closed learning loop that realizes the mutual enhancement of semi-supervised learning and active learning. Furthermore, an adaptive node adjustment method is proposed, which adjusts only a few neurons to adapt to local changes. Experiments on 17 synthetic and real-world datasets show that the proposed approach outperforms other state-of-the-art methods under various labeling budgets.