We are carrying out methodological research for modelling highly complex data which arise from health care and finance (among others: panel studies for economic and natural phenomena, social network, information retrieval and advising systems, and supermarket transactions). This is a challenging aspect of current mainstream research topics in the Science of Big Data. The essential and fundamental part of such research is to find an effective way to reduce the number of dimensions and parameters. The urge for doing so is more pertinent in this age of information expansion as people often access time series data with hundreds or more dimensions.

This project aims to propose new methodology for dimension reduction and to build models with a small number of parameters for ultra-high dimensional time series data (or more generally, dynamic spatio-temporal data). The corresponding theory will be built, and applications of the new methodology will be explored. One of the key problems is how to depict the cross-sectional dependence of high-dimensional data, and the other important question is to capture asymmetry in spatio-temporal data. We will propose new models to overcome such difficulties to develop new statistical methodology and apply random field theory to obtain asymptotic inferences of the proposed models. 

The potential uses of the methodology to be developed in this project are very broad. Direct application areas include digital economics, digital health science, environmental science and financial risk management. Approaches in this project can be applied to the analysis of public health data across many different regions, air pollution data across many locations, transaction data of many companies and 4D image data analysis for maxillofacial surgical planning.