A modern class of continuous-time neural networks — neural ODE, showed impressive results in several areas, including time-series modeling problems. From the time-series models interpretability point of view, the latent neural ODE does not have many advantages over the classical ML methods. An alternative approach is the differential equation discovery algorithms, working in the composite machine learning pipelines, that are making the model as interpretable as the underlying differential equations. We investigate time-variable ODE models, in which interpretable differential equations are extended and refined using neural networks to address the non-stationary time series modeling problem.
Голованев Я.С. (науч. рук. Хватов А.А.) Time-variable ODE discovery for non-stationary time series modeling // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2022]. URL: https://kmu.itmo.ru/digests/article/9214