Federated Learning on Multilabel Evolving Data Streams

Multilabel classification in distributed evolving data stream environment presents significant challenges, including addressing distributed concept drifts and label dependencies. In this study, we introduce two novel solutions employing federated learning (FL) problem transformation techniques to tackle these challenges effectively. Our first approach is an error-driven micro-cluster-based learning strategy that adapts micro-clusters to the evolving data distributions, enabling the handling of concept drifts from various client sources. Our second approach utilizes a graph-based method that leverages graph centrality to capture label dependency and correlation in distributed multilabel data streams. Experimental evaluations reveal that our proposed solutions outperform state-of-the-art methods in terms of multilabel classification metrics. This study highlights the potential of FL in overcoming the challenges associated with distributed multilabel data stream classification. The proposed method’s source code can be accessed publicly on GitHub at the following link: (https://github.com/kholam/FedMuL).