The authors propose aims at learning a better model and simultaenously reduce communication overhead in the Federated Learning paradigm. To that end, the authors propose (1) Cluster-wise straggler dropout and (2) Correlation-based node selection to intelligently drop stragglers and select more related nodes in each cluster to be sent through the network.
The main idea is based on some observations on case study. First, they invesitgate the structurability on the dataset. Hopkin statistics is used on on 6 different HAR datasets and they found strong clusterability. That is, there exist cluster structures in the data. Another pivot study shows that Centralized-cluster models yield higher accuracy and less standard devation.
Given a large amount of data, Centralized learning is better in Fig. 12. Trained models contain some information of users, there are some potential privacy concern if information can be recovered from the model.
In FL paradigm, users' models are sent through the network instead of data due to privacy concern. Hopkin statistics to learn clusterability. KL divergence to measure classifiers' similarities.