Abstract

Multi-output Gaussian processes (MOGPs) are recently extended by usingspectral mixture kernel, which enables expressively pattern extrapolation witha strong interpretation. In particular, Multi-Output Spectral Mixture kernel(MOSM) is a recent, powerful state of the art method. However, MOSM cannotreduce to the ordinary spectral mixture kernel (SM) when using a singlechannel. Moreover, when the spectral density of different channels is eithervery close or very far from each other in the frequency domain, MOSM generatesunreasonable scale effects on cross weights which produces an incorrectdescription of the channel correlation structure. In this paper, we tacklethese drawbacks and introduce a principled multi-output convolution spectralmixture kernel (MOCSM) framework. In our framework, we model channeldependencies through convolution of time and phase delayed spectral mixturesbetween different channels. Results of extensive experiments on synthetic andreal datasets demontrate the advantages of MOCSM and its state of the artperformance.

 

Introduction (beta)

None

 

Conclusion (beta)

None