| Author: | Zhou, Ziang |
| Title: | Towards practical graph neural networks : addressing label scarcity, heterogeneity, and inference efficiency |
| Advisors: | Shi, Jieming (COMP) Li, Qing (COMP) |
| Degree: | Ph.D. |
| Year: | 2026 |
| Department: | Department of Computing |
| Pages: | xviii, 150 pages : color illustrations |
| Language: | English |
| Abstract: | Graph Neural Networks (GNNs) have become a cornerstone for learning on graph-structured data, yet their widespread adoption in real-world scenarios is often impeded by significant practical challenges. This thesis aims to bridge the gap between the theoretical power of GNNs and their practical applicability by systematically addressing three critical limitations: data scarcity, structural heterogeneity, and model inefficiency. First, we confront the challenge of data scarcity, where GNNs often falter when extremely few labeled data is available. To overcome this, we introduce a stabilized self-training framework. This method enhances model performance in few-labeled settings by dynamically generating high-quality pseudo-labels in a robust and balanced manner, thereby stabilizing the training process and significantly improving predictive accuracy with limited supervision. Moving from data limitations to structural complexity, we tackle the prevalent issue of heterogeneous graphs, which contain diverse types of nodes and edges. We identify that conventional GNNs often entangle distinct relational semantics during message passing. In response, we propose SlotGAT, a novel architecture employing a slot-based message passing mechanism. By maintaining separate information channels, or slots, for each node type, SlotGAT effectively preserves rich, type-specific semantics, leading to more expressive and accurate graph representations. Finally, to address the challenge of model efficiency and deployment constraints, we present TINED, a novel GNN-to-MLP knowledge distillation framework. Rather than merely matching output distributions, TINED performs a fine-grained, layer-wise transfer of knowledge. It achieves this through two key innovations: Teacher Injection, which directly transplants compatible parameters from the GNN teacher to the MLP student, and Dirichlet Energy Distillation, which ensures the student model emulates the specific smoothing properties of each layer in the teacher. This results in a lightweight, fast MLP that retains the strong performance of its GNN counterpart without the costly message-passing operations. Collectively, these contributions provide a suite of solutions that enhance the robustness, expressiveness, and efficiency of GNNs. By making these models more stable with scarce data, more capable on complex heterogeneous structures, and more feasible for practical deployment, this thesis pushes GNNs towards being more broadly applicable and impactful tools for real-world problems. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14495