Author Type

Graduate Student

Date of Award

Summer 8-4-2026

Document Type

Dissertation

Publication Status

Version of Record

Submission Date

August 2026

Department

Computer and Electrical Engineering and Computer Science

College Granting Degree

College of Engineering and Computer Science

Department Granting Degree

Electrical Engineering and Computer Science

Degree Name

Doctor of Philosophy (PhD)

Thesis/Dissertation Advisor [Chair]

Xingquan Zhu

Abstract

With the tremendous development of graph neural networks, graph learning has become a dominant solution applied to various applications naturally integrated with graph structures, including traffic networks [82], molecule networks [40, 22, 1, 30, 32], social networks [4], etc. While most existing graph learning solutions can handle homogeneous graphs (graphs with a single node type and a single edge type) and homophily graphs (graphs where node labels tend to be the same as their neighbors) for multi-class node classification downstream tasks well, in real world applications, graph structures can be more complex with heterogeneous graphs (graphs with multiple node types and multiple edge types) or heterophilic graphs (graphs where node labels tend to vary from their neighbors) and more complex downstream tasks such as label distribution learning tasks. In order to enable graph learning to capture more complex graph topology and label space semantics, the dissertation mainly focuses on homogeneous and heterogeneous graph label distribution learning tasks (rich label semantics under different graph topology complexities) and tackling oversmoothing to handle heterophilic graph learning ( alleviating a major bottleneck that prevents graph learning on heterophilic graphs). Specifically, the following three problems will be studied: 1) We systematically survey the existing solutions to alleviate the oversmoothing problem that hinders heterophilic graph learning and present a new measure and a new framework that alleviates the oversmoothing problem; 2) We formulate a novel research problem homogeneous graph label distribution learning with label distribution learning put on homogeneous graph setting and present a new dual-GNN architecture to capture both label-label relation and node-node relation to reflect the more complex relations in the label space. 3) We extend the graph label distribution learning problem from homogeneous graph to heterogeneous graph setting and propose a scalable and efficient heterogeneous graph learning framework that unifies the different heterogeneous meta-paths before graph convolution layers to achieve a more efficient and accurate graph label distribution learning. Finally, For the heterophilic graph learning, we evaluated our proposed model and measure against existing SOTA baselines on various real-world datasets and benchmarks. For the newly formulated graph label distribution learning tasks, we construct new benchmark datasets from existing real-world datasets and evaluated the effectiveness of our proposed frameworks with real-world datasets and benchmarks under standard label distribution learning evaluation metrics.

Download

Share

COinS