Deep Learning for Structured Data: Weak Supervision and Interpretability

Open Access
- Author:
- Zhao, Tianxiang
- Graduate Program:
- Informatics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 31, 2024
- Committee Members:
- Dongwon Lee, Professor in Charge/Director of Graduate Studies
Sharon Huang, Major Field Member
Xiang Zhang, Co-Chair & Dissertation Advisor
Suhang Wang, Co-Chair & Dissertation Advisor
Wang-Chien Lee, Outside Unit & Field Member
Vasant Honavar, Major Field Member - Keywords:
- Weak Supervision
Graph Neural Network
Data Mining
Machine Learning
Sequential Decision Process - Abstract:
- Modern machine learning excels at modeling statistical associations and distributions of observational data, showing strong performance across many tasks. However, the application of ML algorithms faces two distinct challenges: (1) in many real-world data-driven applications, collecting sufficient amount of high-quality labeled data remains as a bottleneck, which is challenging and expensive. Particularly, the widely adopted crowd-sourced data collection pipeline inevitably involves human labelers of varying expertise, which further complicate the label quality. (2) One another challenge is their lack of interpretability. Deep models are known as black boxes, which hinders practitioners' trust in applying them to high-stack applications like healthcare or finances. Motivated by these two problems, in this dissertation, I focus on the improvement of deep learning from these two directions, weakly-supervised learning and model interpretability. Particularly, I focus on structured data, including (1) relational data (graphs), which is a powerful tool to depict non-Euclidean data forms using nodes representing entities and edges modeling relations, and (2) sequential decision processes, which contains a sequence of state-action pairs. Both data forms exist pervasively in many real-world applications, like social networks, protein structures, autonomous driving, etc. In this dissertation, I will introduce some of my representative works addressing these two challenges. In the first two works, I will introduce the extension of GNNs for semi-supervised learning and imbalanced labels. Then, I will present a self-supervised disambiguated learning for more discriminative representation learning on graphs. Followed by them, I will present two works in learning sequential decision-making agents. First is how to imitate decision-making skills from human demonstrations of varying qualities by discovering a set of skills, with each skill to model a different action primitive. A well-performed agent can be obtained by composing those skills hierarchically. Last, I will discuss the strategy of designing a more interpretable neural agent, which can explicitly present its learned knowledge in the form of causal graphs.