Multi-source cross-city transfer learning for region-based spatial distribution prediction

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


PhD Thesis Defence


Title: "Multi-source cross-city transfer learning for region-based spatial 
distribution prediction"

By

Mr. Xu GENG


Abstract:

Deep learning-based spatial map prediction techniques have been widely deployed 
to support many smart city applications. However, the development of these 
techniques is usually hindered by the data scarcity problem due to privacy 
concerns, data regulations, digitization techniques, etc. In this thesis, we 
propose and design a multi-source transfer learning algorithm to solve this 
problem by analyzing, learning, and transferring knowledge from many cities 
with large sizes of available datasets to those cities with limited sizes of 
datasets. By incorporating many source domain cities with high diversity, our 
model is more promising to receive and learn from knowledge with higher 
transferability, quality, and quantity. However, the enlargement of source 
domains brought more challenges to the learning algorithm. For example, the 
source domain pre-trained model is requested to handle data from different 
domains with different sizes, dimensions, complexities, etc. The target domain 
models should be capable of eliminating unuseful information to alleviate the 
potential negative transfer learning problem.

In order to achieve this, we designed the meta-graph, which is a centralized, 
unified, aligned, down-sampled and recoverable graph representation structure 
shared across all domains to store the most general and transferrable instances 
from source domains. We further designed the fractal network to search for the 
best model architecture to adapt to domains with different problem complexities 
given meta-graphs from different domains. Prior to this, we investigated the 
feasibility of utilizing graph structure for learning the spatial patterns in 
urban data, as well as the matching-based cross-city transfer learning 
algorithms as precedent preliminary research. The experiment results have shown 
the proficiency of our proposed algorithm in terms of effectiveness and 
efficiency, compared with the naive and state-of-the-art baselines. Besides, we 
also explored the interpretability, and robustness of the model, as well as its 
extensiveness to related scenarios.


Date:                   Wednesday, 21 August 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 4475
                        Lifts 25/26

Chairman:               Prof. Yingying LI (ISOM)

Committee Members:      Prof. Qiang YANG (Supervisor)
                        Prof. Kai CHEN (Supervisor)
                        Prof. Qiong LUO
                        Prof. Xiaofang ZHOU
                        Prof. Hai YANG (CIVL)
                        Dr. yuanqing ZHENG (PolyU)