Demystifying Deep Learning in Predictive Spatiotemporal Analytics: An Information-Theoretic Framework.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep learning excels in predictive spatiotemporal analytics (PSTA) tasks like disease and traffic prediction.
• Determining optimal deep learning model configurations and understanding their learning capacity for PSTA remains challenging.

Purpose of the Study:

• To provide a comprehensive framework for deep learning model design and information-theoretic analysis in PSTA.
• To demystify the power of deep learning for PSTA in a theoretically sound and explainable manner.

Main Methods:

• Developed a novel interactively and integratively connected deep recurrent neural network (I²DRNN) model with input, hidden, and output modules.
• Employed information-theoretic analysis to examine the information-based learning capacity (i-CAP) of the I²DRNN model.
• Validated the model and its i-CAP through experiments on synthetic and real-world PSTA tasks.

Main Results:

• The I²DRNN model demonstrated superior performance compared to classical and state-of-the-art models across various PSTA tasks.
• The model successfully captured meaningful multiscale spatiotemporal dependencies relevant to real-world scenarios.
• Optimal model configurations, determined by performance, aligned with the theoretically derived necessary and sufficient configurations from i-CAP analysis.

Conclusions:

• The proposed I²DRNN framework offers a theoretically sound and explainable approach to deep learning for PSTA.
• Information-theoretic analysis (i-CAP) provides a method to determine optimal deep learning model configurations for PSTA tasks.
• The I²DRNN model effectively addresses the challenge of learning multiscale spatiotemporal dependencies in complex PSTA applications.