Virtual concept drift detection and adaptation in federated data stream learning

This paper aims to provide valuable insights and solutions for addressing virtual concept drift in federated learning (FL). With the widespread application of FL across various domains, learning from data streams and addressing concept drift (i.e., distributional changes in data) have emerged as a crucial research focus in this field. As a major type of concept drift, virtual concept drift, however, has received very little attention so far. This study provides a deep understanding of how virtual drift in streaming data can affect FL models and how it can be tackled effectively. Firstly, we characterize virtual drift with five temporal and four spatial features and design 21 virtual drifting scenarios to investigate its impact. Then we propose a FL framework handling virtual drift, called FL-HVD. It integrates two key components—virtual drift detection and adaptation. We explore how to detect virtual drift effectively in FL by comparing 11 distribution-based and unsupervised drift detection techniques designed for the traditional data stream learning. In particular, we propose FELK that customize and improve Ks_2samp with exceptional detection accuracy and timeliness. With the alert from the detector, we propose a new model adaptation method FDM, to mitigate the negative impact of the detected virtual drift in FL. The experimental results show that FL-HVD performs the best in accuracy in both synthetic and real-world scenarios.