Probabilistic structure discovery in time series data

TL;DR

This paper introduces a fully Bayesian method for discovering structures in time series data using Gaussian processes, which better captures model uncertainty compared to greedy optimization approaches.

Contribution

The paper proposes a novel Bayesian framework for structure discovery in Gaussian process models, moving beyond greedy methods to infer a full posterior over structures.

Findings

More reliable uncertainty estimates in time series modeling

Improved interpretability of Gaussian process structures

Enhanced flexibility in structure learning

Abstract

Existing methods for structure discovery in time series data construct interpretable, compositional kernels for Gaussian process regression models. While the learned Gaussian process model provides posterior mean and variance estimates, typically the structure is learned via a greedy optimization procedure. This restricts the space of possible solutions and leads to over-confident uncertainty estimates. We introduce a fully Bayesian approach, inferring a full posterior over structures, which more reliably captures the uncertainty of the model.