LatTe Flows: Latent Temporal Flows for Multivariate Sequence Analysis

Magda Amiridi, Gregory Darnell, Sean Jewell

Keywords: time-series analysis, multivariate time series forecasting, latent space, autoregressive models

Abstract: We introduce Latent Temporal Flows (\emph{LatTe-Flows}), a method for probabilistic multivariate time-series analysis tailored for high dimensional systems whose temporal dynamics are driven by variations in a lower-dimensional discriminative subspace. We perform indirect learning from hidden traits of observed sequences by assuming that the random vector representing the data is generated from an unobserved low-dimensional latent vector. \emph{LatTe-Flows} jointly learns auto-encoder mappings to a latent space and learns the temporal distribution of lower-dimensional embeddings of input sequences. Since encoder networks retain only the essential information to generate a latent manifold, the temporal distribution transitions can be more efficiently uncovered by time conditioned Normalizing Flows. The learned latent effects can then be directly transferred into the observed space through the decoder network. We demonstrate that the proposed method significantly outperforms the state-of-the-art on multi-step forecasting benchmarks, while enjoying reduced computational complexity on several real-world datasets. We apply {\emph{LatTe-Flows}} to a challenging sensor-signal forecasting task, using multivariate time-series measurements collected by wearable devices, an increasingly relevant health application.