Latency-Aware Adaptive Learning for Nonstationary Data Streams: Coordinating Concept Drift and Delayed Supervision

Online learning in nonstationary environments is essential for data-stream applications, yet its reliability is often compromised by the coexistence of concept drift and verification latency. Delayed labels may introduce temporal inconsistency in model updates, while evolving distributions require timely adaptation. To address these coupled challenges, an anchor-based temporal consistency management strategy, termed ATCM, is proposed to organize delayed streams into temporally consistent sub-streams through adaptive model allocation and reuse. A drift-guided ensemble of randomized neural networks, termed DGE-RNNets, is further developed to enhance robustness via adaptive aggregation and drift-triggered refreshing. Theoretical regret analysis is provided to characterize the effect of delay patterns on long-term performance. Experiments on synthetic and real-world benchmark streams are conducted under gradual, incremental, and naturally evolving distributional changes. The results show that the proposed framework achieves competitive performance against state-of-the-art methods under the evaluated drift and verification-latency settings. In addition, the proposed framework maintains stronger robustness under increasing delay ratios while providing a balance between predictive performance and computational cost.