Structured Kernel Interpolation for Scalable Mixture of Gaussian Processes
Yu Peng Guo, Ning Xu, Zhenzhou Jin · 2022 7th International Conference on Intelligent Computing and Signal Processing (ICSP) · 2022
Mixture of Gaussian Processes (MoGP) is a promising method that achieves highly accuracy for regression. However, the tremendous training cost of MoGP remains a challenging problem. Although the state-of-the-art sparse MoGP can reduce computational complexity to some extent, the prediction accuracy has been degraded dramatically. Recently, a new structured kernel interpolation (SKI) framework was proposed in the domain of GPs, which generalizes inducing point methods for scalable Gaussian process inference. Based on this amazing idea, we have moved one step further to introduce the SKI framework into MoGP for the sake of addressing the problem mentioned above. In this paper, we propose a scalable mixture of Gaussian processes model (MoKISS-GP), where experts and gating functions are modeled by KISS-GP. Using the Kronecker structure of grid data, the approximate kernel generated by SKI allows us to perform fast calculations while retaining considerable computational accuracy. Finally, experimental results on several well-known datasets demonstrate the effectiveness of our method.