Spatial-Temporal Data Mining with Smart City Applications

PhD Thesis Proposal Defence


Title: "Spatial-Temporal Data Mining with Smart City Applications"

by

Mr. Guanyao LI


Abstract:

Spatial-temporal data are data with both location and time dimensions, such as 
user trajectory, vehicular traffic, bike rental, etc. Massive amount of 
spatial-temporal data have been collected in various sectors of industry. They 
not only reveal user mobility, preference and social activity, but also shed 
insights on the exchange of commodities or assets between regions. Mining 
spatial-temporal data hence plays a pivotal role in enabling many smart city 
applications, such as urban planning, traffic management, logistic and 
supply-chain management, etc.

In this thesis, we first study the problem of assessing spatial-temporal 
similarity for trajectories. We propose a novel and effective measure termed 
STS which considers location noise, and heterogeneous and diverse data sampling 
rates in the data. By estimating the location probability distribution of 
objects at any arbitrary time, STS estimates trajectory similarity with much 
better accuracy as compared with the state-of-the-art approaches.

In an epidemic (e.g., Covid-19), spatial-temporally tracing these close 
contacts as soon as possible is of paramount importance. We next study 
automatic contact tracing when the virus has a lifespan. Leveraging the 
ubiquity of WiFi signals, we propose vContact, a novel, private, and effective 
IoT contact tracing solution given virus lifespan. vContact can be used to 
detect both direct face-to-face and indirect environmental contact. It employs 
novel data processing approaches and an effective WiFi similarity measure and 
achieves better detection performance than other existing approaches in both 
indoor and outdoor scenarios.

Moreover, we study the problem of demand and supply prediction for docked 
bikes. Given station locations and historical rental data of bike flow from 
time 0 to t - 1, we aim to predict the demand and supply at any station at time 
t We propose STGNN-DJD, a novel spatial-temporal graph neural network to 
capture the dynamic and joint spatial-temporal dependency between stations. 
STGNN-DJD employs novel and effective spatial-temporal graphs and aggregators 
to learn the dependency between stations in terms of direct flow between 
stations as well as the correlation of their flow time-series. Compared with 
the state-of-the-art approaches, STGNN DJD achieves significant improvement on 
RMSE and MAE (by 20% ∼50%).

Finally, we study the problem of traffic forecasting without flow data between 
regions. Given the aggregated inflow and outflow traffic of regions over the 
past time slots, we predict the traffic (i.e., aggregated inflow and outflow) 
at the next slot for any region. We propose ST-TIS, a novel, small, and 
accurate spatial-temporal transformer with information infusion and region 
sampling for traffic forecasting. ST-TIS captures the joint and dynamic 
spatial-temporal dependency between regions. It addresses the long-tail issue 
of the canonical transformer and boosts its efficiency. Extensive experiments 
show that ST-TIS is significantly more efficient and accurate than the 
state-of-the-art approaches, with an average improvement of up to 11% on RMSE, 
14% on MAPE, and a reduction of up to 90% on training time and network 
parameters.


Date:			Thursday, 27 January 2022

Time:                  	10:00am - 12:00noon

Zoom Meeting: 
https://hkust.zoom.us/j/97433449997?pwd=U0VtMHNYVmhEMVFkY1pKQm5GZzE0Zz09

Committee Members:	Prof. Gary Chan (Supervisor)
 			Prof. Xiaofang Zhou (Chairperson)
 			Prof. Qiong Luo
 			Prof. Raymond Wong


**** ALL are Welcome ****