Cross-city deep transfer learning on graph representation of region-based spatiotemporal data

PhD Thesis Proposal Defence


Title: "Cross-city deep transfer learning on graph representation 
of region-based spatiotemporal data"

by

Mr. Xu GENG


Abstract:

With the rapid development of urban sensor techniques, spatiotemporal data 
of large volume, variety and velocity has been widely collected, 
processed, warehoused and utilized. The real-time and future distribution 
of the spatiotemporal data has been an important indicator for urban 
planners, governors, commercials and citizens to show the city status, 
interpolate urban phenomenons and make decisions. Machine learning-based 
algorithms, especially deep learning-based algorithms which emerge over 
the past decades in this research field, are the key solutions to this 
problem from the technical perspective. However, the training of deep 
neural networks essentially requires a large amount of labelled data, 
which is usually not available in under-developed cities, new start-ups or 
new markets. This data limitation has become a big barrier to deploying 
accurate algorithms.

This proposal illustrates a top-down approach to solving the data scarcity 
problem, in order to provide accurate, robust and efficient deep learning 
solutions to build spatiotemporal prediction algorithms in data-poor 
cities. First, we propose a graph-based representation for the 
spatiotemporal data and use graph neural networks to solve the supervised 
learning problem in data-rich cities. It encodes various kinds of 
region-wise relationships as the graph structure. The graph-based model is 
expected to be more flexible and accurate. Second, we propose an 
inter-graph transfer learning technique from one source domain city with 
abundant data to a target domain city with limited data. The transfer 
learning algorithm is designed to discover the most similar region pairs 
between a pair of domains and transfer the knowledge through this 
matching. Third, we propose a meta-learning algorithm to discover the 
meta-knowledge from many source domains and construct the target domain 
knowledge from it. The meta-learning algorithm will produce a centralized 
and unified meta-representation to store the most general knowledge from a 
wide range of source domain cities. The target domain with different map 
sizes and data regularities could construct a machine learning model from 
the meta-representation using a limited amount of data with limited 
computation effort and time consumption. Ahead of these technical steps, 
we propose to construct the first multi-city spatiotemporal dataset from a 
wide range of data sources and diversity of cities to enable this research 
plan.

This proposal will provide a detailed research plan for the above-proposed 
research topics, including the intuition, algorithm, experiment design and 
explanation of the methodology. As a background, we will scientifically 
define the problem and introduce the current progress of the related 
research. The aim of this proposal is to enable the proposed research 
design and motivate the experiment investigation.


Date:			Thursday, 23 February 2023

Time:                  	4:30am - 6:30pm

Venue:			Room 5501
 			lifts 25/26

Committee Members:	Prof. Qiang Yang (Supervisor)
 			Prof. Lei Chen (Supervisor)
  			Prof. Kai Chen (Chairperson)
 			Dr. Xiaojuan Ma


**** ALL are Welcome ****