Selective Multi-Source, Non-IID Transfer Learning for Urban Spatio-Temporal Forecasting

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Selective Multi-Source, Non-IID Transfer Learning for Urban 
Spatio-Temporal Forecasting"

By

Mr. Yilun JIN


Abstract:

Spatio-temporal Forecasting (STF) is a fundamental problem in urban 
computing supporting many downstream tasks like transportation, energy, 
meteorology, etc. In recent years, deep learning has achieved tremendous 
success in urban STF. However, the urbanization process creates constant 
needs for new cities, new urban services, and expansion of existing 
services. In these scenarios, the problem of data scarcity negatively 
impacts the accuracy of urban STF and their downstream applications.

Transfer learning is effective in addressing the data scarcity issue in 
computer vision, natural language processing, and recommendation. However, 
spatial and temporal heterogeneity in urban spatio-temporal data poses 
challenges to applying transfer learning for urban STF. Due to such 
heterogeneity, the source city (or cities) contains multiple non-IID latent 
domains, resulting in multi-source, non-IID transfer learning problems. 
Consequently, it is necessary that some latent source domains are helpful to 
the target city, while others may be irrelevant or harmful, which 
compromises the performance of transfer learning for urban STF.

In this thesis, we focus on the problem of selective multi-source, non-IID 
transfer learning for urban STF, whose goal is to select helpful latent 
domains from the source city and transfer the knowledge to the target city. 
We first define the problem of selective transfer learning for urban STF and 
identify three practical sub-problems: selective transfer learning for 
few-shot grid-like urban STF, few-shot graph-like urban STF, and zero-shot 
urban STF. We then present three selective transfer learning frameworks to 
address the sub-problems respectively. For the first sub-problem of 
selective transfer learning for few-shot, grid-like urban STF, we propose 
CrossTReS which selects latent source domains by adaptively re-weighting 
source spatial units, i.e. assigning helpful source spatial units with high 
weights and vice versa. For the second sub-problem of selective transfer 
learning for few-shot, graph-like urban STF, we propose TransGTR that 
selects graph-structured latent source domains with both spatial units and 
spatial relations between them. TransGTR achieves this by learning 
transferable graph structures for the selected spatial units in the source 
city and the target city. For the third sub-problem of selective transfer 
learning for zero-shot transfer learning, we propose KPD-MoE that selects 
latent source domains at test time. KPD-MoE is based on the 
mixture-of-experts architecture which decouples and learns from different 
latent source domains with different experts. During testing, experts which 
learn from most helpful latent source domains are selected for each target 
sample. Extensive experimental results on real-world urban datasets show the 
effectiveness of CrossTReS, TransGTR, and KPD-MoE compared to existing 
non-selective baselines.


Date:                   Wednesday, 16 April 2025

Time:                   10:00am - 12:00noon

Venue:                  Room 4475
                        Lifts 25/26

Chairman:               Prof. Furong GAO (CBE)

Committee Members:      Prof. Qiang YANG (Supervisor)
                        Prof. Kai CHEN (Supervisor)
                        Dr. Yangqiu SONG
                        Dr. Long CHEN
                        Dr. Can YANG (MATH)
                        Prof. Hongxia YANG (PolyU)