Applying Deep Learning to Domain Specific Problems

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


PhD Thesis Defence


Title: "Applying Deep Learning to Domain Specific Problems"

By

Mr. Haoren ZHU


Abstract:

General deep learning models have shown remarkable capability to capture 
intricate patterns across various data types, leading to high accuracy 
across various prediction tasks. However, when these models are applied to 
specific domains, they may encounter challenges related to data scarcity, 
unique characteristics, and distinct constraints. Data scarcity may 
constrain model generalization, increase overfitting risks, and complicate 
pattern extraction from limited datasets. Insensitivity to unique 
domain-specific characteristics can result in the oversight of key features 
crucial for the task. Distinct constraints may influence the design of 
objective functions and evaluation frameworks, thereby shaping the model's 
performance and results. These challenges can restrict the efficacy of 
general deep learning models in addressing domain-specific problems, 
highlighting the need for domain adaptation techniques and tailored model 
designs to enhance performance in specialized contexts.

This thesis explores the application of deep learning methodologies across a 
spectrum of domain-specific subjects, including asset dependency 
forecasting, volatility modeling, breast cancer early detection, and 
influential recommender systems. In the first study, we propose innovative 
transformation techniques for market segmentation within the asset 
dependency matrix, enabling effective representation learning from volatile 
financial data. Additionally, we incorporate a Mix-of-Experts (MoE) 
architecture to address regime-switching dynamics.  The second study bridges 
the gap between stochastic and neural network volatility modeling approaches 
by establishing an equivalence relationship between Generalized 
Autoregressive Conditional Heteroskedasticity (GARCH) models and their 
corresponding neural network counterparts. By integrating the GARCH 
structure into the neural network architecture, our framework effectively 
captures the inherent stylized facts of GARCH models and improves accuracy. 
In the third study, we propose a household solution for early breast cancer 
diagnosis utilizing wearable sensors and time series classification 
techniques. We develop novel noise-filtering and transformation pipelines to 
uncover highly domain-specific patterns from restricted and noisy clinical 
data. The fourth study introduces a novel recommendation paradigm that 
incorporates influential behavior to proactively guide a user's interests 
through a meticulously selected items sequence. We present three tailored 
frameworks adapted from existing recommender systems and develop a unique 
offline evaluation framework and metrics for this purpose. Overall, this 
thesis advances knowledge in deep learning research and provides practical 
insights for professionals and researchers seeking to apply general deep 
learning techniques to solve domain-specific problems.


Date:                   Monday, 24 March 2025

Time:                   4:00pm - 6:00pm

Venue:                  Room 4472
                        Lifts 25/26

Chairman:               Dr. Hai ZHANG (MATH)

Committee Members:      Dr. Wilfred NG (Supervisor)
                        Dr. Pengfei ZHAO (Co-supervisor, BNU-HKBU UIC)
                        Prof. Cunsheng DING
                        Prof. Raymond WONG
                        Prof. Weichuan YU (ECE)
                        Dr. Eric LO (CUHK)