The Automatic Statistician: A Relational Perspective

Abstract

Gaussian Processes (GPs) provide a general and analytically tractable way of capturing complex time-varying, nonparametric functions. The time varying parameters of GPs can be explained as a composition of base kernels such as linear, smoothness or periodicity in that covariance kernels are closed under addition and multiplication. The Automatic Bayesian Covariance Discovery (ABCD) system constructs natural-language description of time-series data by treating unknown time-series data nonparametrically using GPs. Unfortunately, learning a composite covariance kernel with a single time-series dataset often results in less informative kernels instead of finding qualitative distinct descriptions. We address this issue by proposing a relational kernel learning which can model relationship between sets of data and find shared structure among the time series datasets. We show the shared structure can help learning more accurate models for sets of regression problems with some synthetic data, US top market capitalization stock data and US house sales index data.