[1Yin-B] Poster 2

[1Yin-B-09] Development of a General-Purpose Generative Model for Irregular Time Series Using Ornstein–Uhlenbeck Processes and Normalizing Flows

〇Taiki Morinaga Morinaga¹ (1. Hitachi, Ltd.)

Keywords:

time series,generative model,normalizing flow

Time series generative models are used not only for tasks that generate entire sequences, such as video generation, but also for forecasting or imputing values at unobserved time points from partially observed data. However, many existing methods specialize in a particular task, which makes it difficult to handle generation, forecasting, and imputation in a unified manner with a single trained model. Moreover, real world time series data are not necessarily observed at regular intervals, and frameworks that naturally handle irregular sampling are important for many applications. In this study, we propose OUFlow, a new time series generative model that enables generation, forecasting, and imputation within a single model for irregularly sampled time series. OUFlow assumes that latent variables follow a mixture distribution of Ornstein-Uhlenbeck processes and decodes them into the target space via linear transformations and normalizing flows. Owing to the analytical tractability of each module, the posterior distribution of the target variables at arbitrary time points can be derived analytically given observations at an arbitrary set of time points, enabling flexible conditional generation. Compared with existing methods that pursue similar flexibility, the proposed approach achieves state of the art performance on multiple datasets across all tasks.