Learning Non-stationary and Chaotic Behaviors for Spatiotemporal Prediction: An Information Tracking Approach

The objective of the proposed project will be to develop, analyze, and validate a novel learning approach, namely the information tracking neural network (ITNN), for the modeling and characterization of the non-stationary and chaotic behaviors inherent to spatiotemporal data, thereby improving spatiotemporal predictions.

Spatiotemporal prediction is pivotal in numerous fields, such as public health and environmental science, as it enables the precise tracking of changes in pertinent variables or factors across geographical locations and time, thereby informing the decision-making process in a multitude of applications, from controlling infectious diseases to executing effective climate change mitigation strategies.

Although existing machine learning techniques, particularly deep learning techniques, have made notable strides in spatiotemporal prediction tasks, most real-world spatiotemporal data exhibit non-stationary or chaotic behaviors. These data patterns and distributions evolve over time and are extremely sensitive to even minor perturbations, presenting challenges to modeling using current methods. This leads to a key research question: How can we equip prediction models with robustness and adaptability to effectively handle the complex, non-stationary, and chaotic characteristics inherent to spatiotemporal data?

To answer this question, we will devise a novel approach that we call the ITNN. Specifically, we will systematically answer three fundamental questions. (1) On method: How should we quantitatively characterize non-stationary and chaotic behaviors, thereby improving spatiotemporal predictions? (2) On theory: How should we theoretically guarantee the learning performance of the ITNN from the perspective of complex systems modeling? (3) On application: How should we tailor our methodological design of the ITNN to application domains?

To answer the first question, we will develop a deep architecture with an information tracking mechanism to trace, represent, and model the dynamical changes of data. To answer the second question, we will analyze and derive the conditions required for the algorithm stability and solution optimality of the ITNN using important metrics in complex systems modeling. To answer the third question, we will adapt and apply the ITNN to real-world applications by accounting for domain-specific knowledge and data properties.

The ITNN approach has broad implications for machine learning and data analytics. The insights that can be gained adopting this approach are expected to directly impact complex real-world applications such as infectious disease control and climate change mitigation. Furthermore, the proposed project will cultivate local talent by providing training and research opportunities in the field.