A Machine Learning Based Computing Node Selection Algorithm

Abstract

Under the trend of multiple big data technologies, the development of multi-node collaborative parallel computing has made great advances. Limited by data security and data permissions, federated learning has become the current mainstream method to solve the difficult situation. However, for the huge number of computing nodes based on federated learning, there are many redundancies in client nodes. Therefore, the computing resource utilization and model training efficiency can be negatively affected. In this paper, in order to improve the computing efficiency in the parallel computing scenarios such as federated learning, we study a node selection algorithm based on machine learning. In the proposed algorithm, each node implements independent training based on the current global model. According to the contribution to model training accuracy, we select the computing nodes which will be aggregated in central sever afterwards by means of LASSO(Least Absolute Shrinkage and Selection Operator). Additionally, in the experimental simulation part, we add the standard federated averaging (FedAvg) method with random selection as comparison result. Experimental simulation shows that the combination of nodes selected by our proposed algorithm can achieve higher calculation efficiency, and improve the accuracy of federated learning and the mean square error of the accuracy.