Coded Distributed Gaussian Process Regression

Abstract

In this letter, we propose a coded load balancing method for distributed Gaussian process regression over heterogeneous wireless networks, where users with diverse computational and communications capabilities may offload excessive training data onto a computationally stronger central server to reduce collaborative processing times. The offloaded data are transformed using random Fourier feature mapping and encoded with a random orthogonal matrix to prevent transmission of raw data. The proposed method is particularly applicable to compute-intensive applications, where users operate with large datasets.