Time-aware neural ordinary differential equations for incomplete time series modeling.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Internet of Things (IoT) generates vast amounts of time-series data.
• Real-world time-series data frequently contains missing values due to sensor issues or noise.
• Current methods for handling missing time-series data often degrade temporal information and introduce errors.

Purpose of the Study:

• To propose a novel continuous neural network architecture, Time-aware Neural-Ordinary Differential Equations (TN-ODE), for modeling incomplete time-series data.
• To enable accurate imputation of missing values at any time point.
• To facilitate multi-step predictions at specified future time points.

Main Methods:

• Developed TN-ODE, a continuous neural network architecture.
• Utilized a time-aware Long Short-Term Memory (LSTM) as an encoder to learn from partial observations.
• Parameterized the derivative of latent states using a fully connected network for continuous-time dynamics.

Main Results:

• TN-ODE demonstrated superior performance in data interpolation and extrapolation tasks compared to baseline methods.
• The model achieved lower Mean Square Error (MSE) for imputation and prediction.
• TN-ODE also showed improved accuracy in downstream time-series classification tasks.

Conclusions:

• TN-ODE offers an effective solution for modeling incomplete time-series data without compromising temporal information.
• The architecture successfully handles missing value imputation and multi-step prediction.
• TN-ODE represents a significant advancement in time-series analysis for IoT applications.