Latent Network Construction for Univariate Time Series Based on Variational Auto-Encode

Area of Science:

• Data Science
• Network Science
• Time Series Analysis

Background:

• Time series analysis is crucial for information processing.
• Converting time series to complex networks offers new analytical perspectives.
• Existing methods may not fully capture the underlying structure of univariate time series.

Purpose of the Study:

• To explore the construction of latent networks for univariate time series.
• To utilize Variational Auto-Encoder (VAE) for latent network generation.
• To develop a novel data structure for enhanced time series analysis.

Main Methods:

• Trained a Variational Auto-Encoder (VAE) to generate latent probability distributions.
• Decomposed multivariate Gaussian distributions into univariate Gaussian distributions.
• Measured distances between univariate Gaussian distributions on a statistical manifold for network construction.

Main Results:

• Successfully constructed latent networks from univariate time series.
• Demonstrated that the latent network effectively retains original time series information.
• The proposed latent network serves as a valuable new data structure.

Conclusions:

• The VAE-based latent network construction is a viable method for time series analysis.
• Latent networks offer a promising new data representation for downstream tasks.
• This approach enhances the understanding and analysis of univariate time series data.