Deep Switching Auto-Regressive Factorization: Application to Time Series Forecasting

We introduce deep switching auto-regressive factorization (DSARF), a generative model for spatio-temporal data with the capability to unravel recurring patterns in and perform robust short- long-term predictions. Similar other factor analysis methods, DSARF approximates high dimensional by product between time dependent weights spatially factors. These factors are turn represented terms of lower latent variables that inferred using stochastic variational inference. is different from state-of-the-art techniques it parameterizes vector likelihood governed Markovian prior, which able capture non-linear inter-dependencies among characterize multimodal temporal dynamics. This results flexible hierarchical can be extended (i) provide collection potentially interpretable states abstracted process dynamics, (ii) series prediction complex multi-relational setting. Our extensive experiments, include simulated real wide range applications such as climate change, weather forecasting, traffic, infectious disease spread nonlinear physical systems attest superior performance long- short-term error, when compared methods.