Distributed Principal Component Analysis Based on Randomized Low-Rank Approximation

Distributed PCA aims to implement dimension reduction for data stored on multiple agents. The conventional distributed PCA encounters the bottleneck of computation when the dimension of 10-cal data is large. In this work, we propose a distributed PCA algorithm with local processing based on randomized methods for the star network topology (master-slave networks) with distributed row observations. Local matrix approximation with randomized methods allows us to accelerate the computation with an acceptable loss of precision significantly. The results of numerical experiments show that the proposed algorithm can achieve satisfactory decomposition results with much lower computational complexity.