Overcoming Data Heterogeneity, Dependency, and Noise: A Novel Spatio-Temporal Learning Framework

The goal of this project is to develop and demonstrate a novel spatio-temporal learning (STL) framework to overcome the fundamental challenges of data heterogeneity, dependency, and noise in a systematic way for spatio-temporal predictive analytics.

Spatio-temporal predictive analytics, which aims to analyze and model data with both spatial and temporal information for future prediction and forecasting, has long attracted extensive research interest and has assumed increasing importance as the massive increases in both data availability and data diversity make spatio-temporal data more ubiquitous than ever. The problem of modeling spatio-temporal data is interesting and important but quite challenging due to (1) substantial data heterogeneity in terms of data formats, sizes, and resolutions; (2) complex data dependency at varying spatio- temporal scales within and/or across heterogenous data sources; and (3) various types of noise in the heterogenous spatio-temporal data.

To tackle these fundamental challenges, in this project, we will develop a novel STL framework that effectively integrates, models, and analyzes the spatio-temporal data for predictive analytics. Under the STL framework, we will further propose three new methods:
(1) An STL method with Tensor representation (referred to as STL-T) to effectively integrate data from heterogeneous sources while preserving the intrinsic spatio- temporal structure of data from each source via tensor-based operations.
(2) An STL-T method with Deep architecture (referred to as STL-TD) to quantitatively characterize, effectively learn, and faithfully interpret the complex dependency of heterogenous spatio-temporal data.
(3) An STL-TD method with Regularizer (referred to as STL-TDR) to address various sources of data noise for robust spatio-temporal predictive analytics.

To investigate these issues in a systematic way, we will (i) formally define and formulate the problems; (ii) theoretically analyze the developed formulations and models; (iii) computationally validate the solutions; and (iv) practically evaluate the proposed framework and methods in three real-world scenarios with domain-specific implementations.

The proposed framework and methods are general in nature and will contribute to both machine learning foundations and data analytics techniques in academia. The results derived from the proposed framework and methods, especially in the era of big data, will have practical benefit in challenging application domains of spatio-temporal predictive analytics, such as epidemiology, transportation, and climate science. This project will nurture local talent in the field by creating training and research opportunities for both undergraduate and graduate students.