Research On Federated Learning Algorithm For Class Imbalance

In the era of big data,the explosive growth of data and the continuous improvement of machine learning technology are driving the development of all walks of life.In order to obtain better training models and more considerable economic benefits,the extensive collection of users’ private data has become an unavoidable problem.How to utilize the value of private data within the scope permitted by law has become more and more urgent.As an emerging paradigm of machine learning,federated learning provides a novel solution for user data sharing,which enables participants to jointly train models through the coordination of a central server without exposing local data.However,in practical applications,the local data of each participant often exhibits non-independent and identical distribution characteristics,and even class imbalance occurs.This will not only affect the convergence speed of the model,but also cause unnecessary communication consumption,and even cause the model to diverge,which brings great challenges to the training of federated learning.Therefore,it is necessary to design a federated learning algorithm that can effectively deal with the above challenges for the unbalanced data in the federated learning field.This paper designs a federated learning client online intelligent selection algorithm for unbalanced class distribution.The algorithm does not need to share the client’s local data set.On the premise of ensuring that the user’s private data does not go out of the local area,the similarity coefficient is used to measure the difference between the local data distribution and the target distribution,thereby proposing a class balance selection algorithm.By actively selecting an appropriate subset of clients to offset the convergence problems caused by data heterogeneity to the model.On this basis,the combined multi-armed bandit model in the field of reinforcement learning is used to balance the development and exploration of client device selection,and further seek the optimal client subset.This enables the algorithm to improve the accuracy and convergence speed of the global model under different types of data heterogeneity.In order to further optimize the performance of the model under globally imbalanced data,this paper improves the general clustering federated learning framework in personalized federated learning.Based on the above work,an adaptive clustering federated learning algorithm for global imbalanced datasets is proposed.The algorithm introduces the ESA framework in the field of privacy protection,and uses the client’s local data distribution vector to guide the server’s clustering selection under the premise that the privacy data is protected,which improves the clustering accuracy.Furthermore,an adaptive clustering federated learning algorithm is proposed,which adaptively determines the number of clusters according to the similarity between the client’s local vectors,and solves the dependence on the optimal number of clusters k in general clustering federated learning problems.