Spatiotemporal information conversion machine for time-series forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Robust time-series forecasting for nonlinear systems using only observed data is challenging.
• Existing methods struggle with high-dimensional and noisy datasets.

Purpose of the Study:

• To develop a novel neural network framework for accurate and robust time-series forecasting.
• To improve forecasting by incorporating causal factor inference.

Main Methods:

• Developed the spatiotemporal information conversion machine (STICM), a neural network framework.
• Employed spatial-temporal information (STI) transformation and temporal convolutional networks.
• Integrated Granger causality to identify and utilize causal factors for improved robustness.

Main Results:

• STICM demonstrated superior and robust performance on benchmark and real-world datasets.
• The framework accurately forecasts time series, even when data is perturbed by noise.
• STICM successfully inferred causal factors, enhancing forecasting accuracy and robustness.

Conclusions:

• STICM offers a powerful, model-free approach for time-series forecasting based solely on observed data.
• The framework has significant potential for practical AI applications and dynamical data exploration.
• STICM provides a novel method for analyzing high-dimensional data in a dynamic manner.