Univariate Time Series Analysis with Hyper Neural ODE
Yevhen V. Koshel, Vasiliy Ye. Belozyorov · Journal of Optimization Differential Equations and their Applications · 2023
Neural ordinary differential equations (NODE) are ordinary differential equations whose right-hand side is determined by a neural network. Hyper NODE (hNODE) is a special type of neural network architecture, which is aimed at creating such NODE system that regulates its own parameters based on known input data. The article uses a new approach to the study of one-dimensional time series, the basis of which is the hNODE system. This system takes into account the relationship between the input data and its latent representation in the network and uses an explicit parametrization when controlling the latent flow. The proposed model is tested on artificial time series of data. The influence of some activation functions (besides sigmoid and hyperbolic tangent) on the quality of the forecast is also considered.