Latent adversarial regularized autoencoder for high-dimensional probabilistic time series prediction.

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Probabilistic time series prediction is crucial for many applications.
• High-dimensional time series present challenges due to inter-series dependencies.
• Existing methods often assume independence, limiting their effectiveness.

Purpose of the Study:

• To propose a novel probabilistic model for high-dimensional multivariate time series prediction.
• To address the limitations of existing methods in capturing cross-series relationships.
• To develop a model capable of modeling complex distributions in multivariate time series data.

Main Methods:

• Introduced TimeLAR (latent adversarial regularized autoencoder), integrating Generative Adversarial Networks (GANs) and autoencoders.
• Utilized autoencoder mapping to learn cross-series relationships and encode them into latent variables.
• Employed a modified Transformer for latent variables to capture temporal patterns and a GAN for performance refinement.

Main Results:

• TimeLAR effectively learns cross-series relationships and global temporal patterns.
• The model successfully infers latent space prediction distributions for multi-step predictions.
• Empirical validation on five real-world datasets demonstrates TimeLAR's effectiveness.

Conclusions:

• TimeLAR offers a powerful new approach for high-dimensional multivariate time series prediction.
• The integration of autoencoders, Transformers, and GANs enables robust probabilistic forecasting.
• The model shows promise in diverse application domains including transportation and electricity.