Probabilistic sequential matrix factorization

Abstract

We introduce the probabilistic sequential matrix factorization (PSMF) method for factorizing time-varying and non-stationary datasets consisting of high-dimensional time-series. In particular, we consider nonlinear Gaussian state-space models where sequential approximate inference results in the factorization of a data matrix into a dictionary and time-varying coefficients with potentially nonlinear Markovian dependencies. The assumed Markovian structure on the coefficients enables us to encode temporal dependencies into a low-dimensional feature space. The proposed inference method is solely based on an approximate extended Kalman filtering scheme, which makes the resulting method particularly efficient. PSMF can account for temporal nonlinearities and, more importantly, can be used to calibrate and estimate generic differentiable nonlinear subspace models. We also introduce a robust version of PSMF, called rPSMF, which uses Student-t filters to handle model misspecification. We show that PSMF can be used in multiple contexts: modeling time series with a periodic subspace, robustifying changepoint detection methods, and imputing missing data in several high-dimensional time-series, such as measurements of pollutants across London.

arXiv commentAccepted for publication at AISTATS 2021

References in corpus (9)

• Optimal detection of changepoints with a linear computational cost
• Bayesian Online Changepoint Detection
• Online EM Algorithm for Latent Data Models
• Learning Dynamical Systems from Partial Observations
• Dictionary Learning for Massive Matrix Factorization
• A Probabilistic Least-Mean-Squares Filter
• A State-Space Approach to Dynamic Nonnegative Matrix Factorization
• Robust Bayesian Filtering and Smoothing Using Student's t Distribution
• Decomposing Temperature Time Series with Non-Negative Matrix Factorization