DYNAFLOW: NEURAL ORDINARY DIFFERENTIAL EQUATIONS FOR IRREGULARLY SAMPLED TIME SERIES

Abstract

Modeling irregularly sampled time series is a fundamental challenge in domains like healthcare and finance, where data is inherently sparse and asynchronous. Traditional deep learning models, such as Recurrent Neural Networks (RNNs), struggle with this data due to their discrete-time nature, often relying on ad-hoc imputation or masking that can bias the model. This paper introduces DynaFlow, a novel continuous-time framework built on Neural Ordinary Differential Equations (Neural ODEs). By parameterizing the hidden state dynamics as an ODE, DynaFlow naturally adapts to irregular sampling, using a numerical solver to integrate information continuously between observations and a gated mechanism to incorporate new data. We evaluate DynaFlow on classification, forecasting, and imputation tasks across several real-world and synthetic datasets. Results demonstrate that our method consistently outperforms state-of-the-art models, including GRU-D and ODE-RNN, in both accuracy and robustness. This advantage is particularly pronounced under high data sparsity. Furthermore, DynaFlow provides the auxiliary benefit of generating smooth, interpretable latent trajectories, offering a more faithful representation of continuous underlying processes. Our work establishes Neural ODEs as a powerful foundation for irregular time series analysis.