Abstract
Collaborative learning such as federated learning enables to train a global prediction model in a distributed way without the need to share the training data. However, most existing schemes adopt deep learning models and require all local models to have the same architecture as the global model, making them unsuitable for applications using resource- and bandwidth-hungry devices. In this paper, we present CloREF, a novel rule-based collaborative learning framework, that allows participating devices to use different local learning models. A rule extraction method is firstly proposed to bridge the heterogeneity of local learning models by approximating their decision boundaries. Then a novel rule fusion and selection mechanism is designed based on evolutionary optimization to integrate the knowledge learned by all local models. Experimental results on a number of synthesized and real-world datasets demonstrate that the rules generated by our rule extraction method can mimic the behaviors of various learning models with high fidelity (>0.95 in most tests), and CloREF gives comparable and sometimes even better AUC compared with the best-performing model trained centrally.